godeps: update leveldb and snappy, dump serpent-go

73 files changed, 7989 insertions, 3091 deletions

diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch.go
index 0d7911eca..ccf390c9c 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch.go
@@ -8,65 +8,84 @@ package leveldb
 
 import (
 	"encoding/binary"
-	"errors"
+	"fmt"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/memdb"
 )
 
-var (
-	errBatchTooShort  = errors.New("leveldb: batch is too short")
-	errBatchBadRecord = errors.New("leveldb: bad record in batch")
-)
+type ErrBatchCorrupted struct {
+	Reason string
+}
+
+func (e *ErrBatchCorrupted) Error() string {
+	return fmt.Sprintf("leveldb: batch corrupted: %s", e.Reason)
+}
+
+func newErrBatchCorrupted(reason string) error {
+	return errors.NewErrCorrupted(nil, &ErrBatchCorrupted{reason})
+}
 
-const kBatchHdrLen = 8 + 4
+const (
+	batchHdrLen  = 8 + 4
+	batchGrowRec = 3000
+)
 
-type batchReplay interface {
-	put(key, value []byte, seq uint64)
-	delete(key []byte, seq uint64)
+type BatchReplay interface {
+	Put(key, value []byte)
+	Delete(key []byte)
 }
 
 // Batch is a write batch.
 type Batch struct {
-	buf        []byte
+	data       []byte
 	rLen, bLen int
 	seq        uint64
 	sync       bool
 }
 
 func (b *Batch) grow(n int) {
-	off := len(b.buf)
+	off := len(b.data)
 	if off == 0 {
-		// include headers
-		off = kBatchHdrLen
-		n += off
+		off = batchHdrLen
+		if b.data != nil {
+			b.data = b.data[:off]
+		}
 	}
-	if cap(b.buf)-off >= n {
-		return
+	if cap(b.data)-off < n {
+		if b.data == nil {
+			b.data = make([]byte, off, off+n)
+		} else {
+			odata := b.data
+			div := 1
+			if b.rLen > batchGrowRec {
+				div = b.rLen / batchGrowRec
+			}
+			b.data = make([]byte, off, off+n+(off-batchHdrLen)/div)
+			copy(b.data, odata)
+		}
 	}
-	buf := make([]byte, 2*cap(b.buf)+n)
-	copy(buf, b.buf)
-	b.buf = buf[:off]
 }
 
-func (b *Batch) appendRec(t vType, key, value []byte) {
+func (b *Batch) appendRec(kt kType, key, value []byte) {
 	n := 1 + binary.MaxVarintLen32 + len(key)
-	if t == tVal {
+	if kt == ktVal {
 		n += binary.MaxVarintLen32 + len(value)
 	}
 	b.grow(n)
-	off := len(b.buf)
-	buf := b.buf[:off+n]
-	buf[off] = byte(t)
+	off := len(b.data)
+	data := b.data[:off+n]
+	data[off] = byte(kt)
 	off += 1
-	off += binary.PutUvarint(buf[off:], uint64(len(key)))
-	copy(buf[off:], key)
+	off += binary.PutUvarint(data[off:], uint64(len(key)))
+	copy(data[off:], key)
 	off += len(key)
-	if t == tVal {
-		off += binary.PutUvarint(buf[off:], uint64(len(value)))
-		copy(buf[off:], value)
+	if kt == ktVal {
+		off += binary.PutUvarint(data[off:], uint64(len(value)))
+		copy(data[off:], value)
 		off += len(value)
 	}
-	b.buf = buf[:off]
+	b.data = data[:off]
 	b.rLen++
 	//  Include 8-byte ikey header
 	b.bLen += len(key) + len(value) + 8
@@ -75,18 +94,51 @@ func (b *Batch) appendRec(t vType, key, value []byte) {
 // Put appends 'put operation' of the given key/value pair to the batch.
 // It is safe to modify the contents of the argument after Put returns.
 func (b *Batch) Put(key, value []byte) {
-	b.appendRec(tVal, key, value)
+	b.appendRec(ktVal, key, value)
 }
 
 // Delete appends 'delete operation' of the given key to the batch.
 // It is safe to modify the contents of the argument after Delete returns.
 func (b *Batch) Delete(key []byte) {
-	b.appendRec(tDel, key, nil)
+	b.appendRec(ktDel, key, nil)
+}
+
+// Dump dumps batch contents. The returned slice can be loaded into the
+// batch using Load method.
+// The returned slice is not its own copy, so the contents should not be
+// modified.
+func (b *Batch) Dump() []byte {
+	return b.encode()
+}
+
+// Load loads given slice into the batch. Previous contents of the batch
+// will be discarded.
+// The given slice will not be copied and will be used as batch buffer, so
+// it is not safe to modify the contents of the slice.
+func (b *Batch) Load(data []byte) error {
+	return b.decode(0, data)
+}
+
+// Replay replays batch contents.
+func (b *Batch) Replay(r BatchReplay) error {
+	return b.decodeRec(func(i int, kt kType, key, value []byte) {
+		switch kt {
+		case ktVal:
+			r.Put(key, value)
+		case ktDel:
+			r.Delete(key)
+		}
+	})
+}
+
+// Len returns number of records in the batch.
+func (b *Batch) Len() int {
+	return b.rLen
 }
 
 // Reset resets the batch.
 func (b *Batch) Reset() {
-	b.buf = nil
+	b.data = b.data[:0]
 	b.seq = 0
 	b.rLen = 0
 	b.bLen = 0
@@ -97,24 +149,10 @@ func (b *Batch) init(sync bool) {
 	b.sync = sync
 }
 
-func (b *Batch) put(key, value []byte, seq uint64) {
-	if b.rLen == 0 {
-		b.seq = seq
-	}
-	b.Put(key, value)
-}
-
-func (b *Batch) delete(key []byte, seq uint64) {
-	if b.rLen == 0 {
-		b.seq = seq
-	}
-	b.Delete(key)
-}
-
 func (b *Batch) append(p *Batch) {
 	if p.rLen > 0 {
-		b.grow(len(p.buf) - kBatchHdrLen)
-		b.buf = append(b.buf, p.buf[kBatchHdrLen:]...)
+		b.grow(len(p.data) - batchHdrLen)
+		b.data = append(b.data, p.data[batchHdrLen:]...)
 		b.rLen += p.rLen
 	}
 	if p.sync {
@@ -122,95 +160,93 @@ func (b *Batch) append(p *Batch) {
 	}
 }
 
-func (b *Batch) len() int {
-	return b.rLen
-}
-
+// size returns sums of key/value pair length plus 8-bytes ikey.
 func (b *Batch) size() int {
 	return b.bLen
 }
 
 func (b *Batch) encode() []byte {
 	b.grow(0)
-	binary.LittleEndian.PutUint64(b.buf, b.seq)
-	binary.LittleEndian.PutUint32(b.buf[8:], uint32(b.rLen))
+	binary.LittleEndian.PutUint64(b.data, b.seq)
+	binary.LittleEndian.PutUint32(b.data[8:], uint32(b.rLen))
 
-	return b.buf
+	return b.data
 }
 
-func (b *Batch) decode(buf []byte) error {
-	if len(buf) < kBatchHdrLen {
-		return errBatchTooShort
+func (b *Batch) decode(prevSeq uint64, data []byte) error {
+	if len(data) < batchHdrLen {
+		return newErrBatchCorrupted("too short")
 	}
 
-	b.seq = binary.LittleEndian.Uint64(buf)
-	b.rLen = int(binary.LittleEndian.Uint32(buf[8:]))
+	b.seq = binary.LittleEndian.Uint64(data)
+	if b.seq < prevSeq {
+		return newErrBatchCorrupted("invalid sequence number")
+	}
+	b.rLen = int(binary.LittleEndian.Uint32(data[8:]))
+	if b.rLen < 0 {
+		return newErrBatchCorrupted("invalid records length")
+	}
 	// No need to be precise at this point, it won't be used anyway
-	b.bLen = len(buf) - kBatchHdrLen
-	b.buf = buf
+	b.bLen = len(data) - batchHdrLen
+	b.data = data
 
 	return nil
 }
 
-func (b *Batch) decodeRec(f func(i int, t vType, key, value []byte)) error {
-	off := kBatchHdrLen
+func (b *Batch) decodeRec(f func(i int, kt kType, key, value []byte)) (err error) {
+	off := batchHdrLen
 	for i := 0; i < b.rLen; i++ {
-		if off >= len(b.buf) {
-			return errors.New("leveldb: invalid batch record length")
+		if off >= len(b.data) {
+			return newErrBatchCorrupted("invalid records length")
 		}
 
-		t := vType(b.buf[off])
-		if t > tVal {
-			return errors.New("leveldb: invalid batch record type in batch")
+		kt := kType(b.data[off])
+		if kt > ktVal {
+			return newErrBatchCorrupted("bad record: invalid type")
 		}
 		off += 1
 
-		x, n := binary.Uvarint(b.buf[off:])
+		x, n := binary.Uvarint(b.data[off:])
 		off += n
-		if n <= 0 || off+int(x) > len(b.buf) {
-			return errBatchBadRecord
+		if n <= 0 || off+int(x) > len(b.data) {
+			return newErrBatchCorrupted("bad record: invalid key length")
 		}
-		key := b.buf[off : off+int(x)]
+		key := b.data[off : off+int(x)]
 		off += int(x)
-
 		var value []byte
-		if t == tVal {
-			x, n := binary.Uvarint(b.buf[off:])
+		if kt == ktVal {
+			x, n := binary.Uvarint(b.data[off:])
 			off += n
-			if n <= 0 || off+int(x) > len(b.buf) {
-				return errBatchBadRecord
+			if n <= 0 || off+int(x) > len(b.data) {
+				return newErrBatchCorrupted("bad record: invalid value length")
 			}
-			value = b.buf[off : off+int(x)]
+			value = b.data[off : off+int(x)]
 			off += int(x)
 		}
 
-		f(i, t, key, value)
+		f(i, kt, key, value)
 	}
 
 	return nil
 }
 
-func (b *Batch) replay(to batchReplay) error {
-	return b.decodeRec(func(i int, t vType, key, value []byte) {
-		switch t {
-		case tVal:
-			to.put(key, value, b.seq+uint64(i))
-		case tDel:
-			to.delete(key, b.seq+uint64(i))
-		}
-	})
-}
-
 func (b *Batch) memReplay(to *memdb.DB) error {
-	return b.decodeRec(func(i int, t vType, key, value []byte) {
-		ikey := newIKey(key, b.seq+uint64(i), t)
+	return b.decodeRec(func(i int, kt kType, key, value []byte) {
+		ikey := newIkey(key, b.seq+uint64(i), kt)
 		to.Put(ikey, value)
 	})
 }
 
+func (b *Batch) memDecodeAndReplay(prevSeq uint64, data []byte, to *memdb.DB) error {
+	if err := b.decode(prevSeq, data); err != nil {
+		return err
+	}
+	return b.memReplay(to)
+}
+
 func (b *Batch) revertMemReplay(to *memdb.DB) error {
-	return b.decodeRec(func(i int, t vType, key, value []byte) {
-		ikey := newIKey(key, b.seq+uint64(i), t)
+	return b.decodeRec(func(i int, kt kType, key, value []byte) {
+		ikey := newIkey(key, b.seq+uint64(i), kt)
 		to.Delete(ikey)
 	})
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch_test.go
index 19b749b8f..7fc842f4f 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/batch_test.go
@@ -15,7 +15,7 @@ import (
 )
 
 type tbRec struct {
-	t          vType
+	kt         kType
 	key, value []byte
 }
 
@@ -23,39 +23,39 @@ type testBatch struct {
 	rec []*tbRec
 }
 
-func (p *testBatch) put(key, value []byte, seq uint64) {
-	p.rec = append(p.rec, &tbRec{tVal, key, value})
+func (p *testBatch) Put(key, value []byte) {
+	p.rec = append(p.rec, &tbRec{ktVal, key, value})
 }
 
-func (p *testBatch) delete(key []byte, seq uint64) {
-	p.rec = append(p.rec, &tbRec{tDel, key, nil})
+func (p *testBatch) Delete(key []byte) {
+	p.rec = append(p.rec, &tbRec{ktDel, key, nil})
 }
 
 func compareBatch(t *testing.T, b1, b2 *Batch) {
 	if b1.seq != b2.seq {
 		t.Errorf("invalid seq number want %d, got %d", b1.seq, b2.seq)
 	}
-	if b1.len() != b2.len() {
-		t.Fatalf("invalid record length want %d, got %d", b1.len(), b2.len())
+	if b1.Len() != b2.Len() {
+		t.Fatalf("invalid record length want %d, got %d", b1.Len(), b2.Len())
 	}
 	p1, p2 := new(testBatch), new(testBatch)
-	err := b1.replay(p1)
+	err := b1.Replay(p1)
 	if err != nil {
 		t.Fatal("error when replaying batch 1: ", err)
 	}
-	err = b2.replay(p2)
+	err = b2.Replay(p2)
 	if err != nil {
 		t.Fatal("error when replaying batch 2: ", err)
 	}
 	for i := range p1.rec {
 		r1, r2 := p1.rec[i], p2.rec[i]
-		if r1.t != r2.t {
-			t.Errorf("invalid type on record '%d' want %d, got %d", i, r1.t, r2.t)
+		if r1.kt != r2.kt {
+			t.Errorf("invalid type on record '%d' want %d, got %d", i, r1.kt, r2.kt)
 		}
 		if !bytes.Equal(r1.key, r2.key) {
 			t.Errorf("invalid key on record '%d' want %s, got %s", i, string(r1.key), string(r2.key))
 		}
-		if r1.t == tVal {
+		if r1.kt == ktVal {
 			if !bytes.Equal(r1.value, r2.value) {
 				t.Errorf("invalid value on record '%d' want %s, got %s", i, string(r1.value), string(r2.value))
 			}
@@ -75,7 +75,7 @@ func TestBatch_EncodeDecode(t *testing.T) {
 	b1.Delete([]byte("k"))
 	buf := b1.encode()
 	b2 := new(Batch)
-	err := b2.decode(buf)
+	err := b2.decode(0, buf)
 	if err != nil {
 		t.Error("error when decoding batch: ", err)
 	}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench2_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench2_test.go
new file mode 100644
index 000000000..0dd60fd82
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench2_test.go
@@ -0,0 +1,58 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build !go1.2
+
+package leveldb
+
+import (
+	"sync/atomic"
+	"testing"
+)
+
+func BenchmarkDBReadConcurrent(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+	defer p.close()
+
+	b.ResetTimer()
+	b.SetBytes(116)
+
+	b.RunParallel(func(pb *testing.PB) {
+		iter := p.newIter()
+		defer iter.Release()
+		for pb.Next() && iter.Next() {
+		}
+	})
+}
+
+func BenchmarkDBReadConcurrent2(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+	defer p.close()
+
+	b.ResetTimer()
+	b.SetBytes(116)
+
+	var dir uint32
+	b.RunParallel(func(pb *testing.PB) {
+		iter := p.newIter()
+		defer iter.Release()
+		if atomic.AddUint32(&dir, 1)%2 == 0 {
+			for pb.Next() && iter.Next() {
+			}
+		} else {
+			if pb.Next() && iter.Last() {
+				for pb.Next() && iter.Prev() {
+				}
+			}
+		}
+	})
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench_test.go
index ea6801a89..91b426709 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/bench_test.go
@@ -170,7 +170,7 @@ func (p *dbBench) writes(perBatch int) {
 	b.SetBytes(116)
 }
 
-func (p *dbBench) drop() {
+func (p *dbBench) gc() {
 	p.keys, p.values = nil, nil
 	runtime.GC()
 }
@@ -249,6 +249,9 @@ func (p *dbBench) newIter() iterator.Iterator {
 }
 
 func (p *dbBench) close() {
+	if bp, err := p.db.GetProperty("leveldb.blockpool"); err == nil {
+		p.b.Log("Block pool stats: ", bp)
+	}
 	p.db.Close()
 	p.stor.Close()
 	os.RemoveAll(benchDB)
@@ -331,7 +334,7 @@ func BenchmarkDBRead(b *testing.B) {
 	p := openDBBench(b, false)
 	p.populate(b.N)
 	p.fill()
-	p.drop()
+	p.gc()
 
 	iter := p.newIter()
 	b.ResetTimer()
@@ -362,7 +365,7 @@ func BenchmarkDBReadUncompressed(b *testing.B) {
 	p := openDBBench(b, true)
 	p.populate(b.N)
 	p.fill()
-	p.drop()
+	p.gc()
 
 	iter := p.newIter()
 	b.ResetTimer()
@@ -379,7 +382,7 @@ func BenchmarkDBReadTable(b *testing.B) {
 	p.populate(b.N)
 	p.fill()
 	p.reopen()
-	p.drop()
+	p.gc()
 
 	iter := p.newIter()
 	b.ResetTimer()
@@ -395,7 +398,7 @@ func BenchmarkDBReadReverse(b *testing.B) {
 	p := openDBBench(b, false)
 	p.populate(b.N)
 	p.fill()
-	p.drop()
+	p.gc()
 
 	iter := p.newIter()
 	b.ResetTimer()
@@ -413,7 +416,7 @@ func BenchmarkDBReadReverseTable(b *testing.B) {
 	p.populate(b.N)
 	p.fill()
 	p.reopen()
-	p.drop()
+	p.gc()
 
 	iter := p.newIter()
 	b.ResetTimer()
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/bench2_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/bench2_test.go
new file mode 100644
index 000000000..175e22203
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/bench2_test.go
@@ -0,0 +1,30 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build !go1.2
+
+package cache
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func BenchmarkLRUCache(b *testing.B) {
+	c := NewCache(NewLRU(10000))
+
+	b.SetParallelism(10)
+	b.RunParallel(func(pb *testing.PB) {
+		r := rand.New(rand.NewSource(time.Now().UnixNano()))
+
+		for pb.Next() {
+			key := uint64(r.Intn(1000000))
+			c.Get(0, key, func() (int, Value) {
+				return 1, key
+			}).Release()
+		}
+	})
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache.go
index 9b6a74977..c9670de5d 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache.go
@@ -8,118 +8,669 @@
 package cache
 
 import (
+	"sync"
 	"sync/atomic"
+	"unsafe"
+
+	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-// SetFunc used by Namespace.Get method to create a cache object. SetFunc
-// may return ok false, in that case the cache object will not be created.
-type SetFunc func() (ok bool, value interface{}, charge int, fin SetFin)
+// Cacher provides interface to implements a caching functionality.
+// An implementation must be goroutine-safe.
+type Cacher interface {
+	// Capacity returns cache capacity.
+	Capacity() int
 
-// SetFin will be called when corresponding cache object are released.
-type SetFin func()
+	// SetCapacity sets cache capacity.
+	SetCapacity(capacity int)
 
-// DelFin will be called when corresponding cache object are released.
-// DelFin will be called after SetFin. The exist is true if the corresponding
-// cache object is actually exist in the cache tree.
-type DelFin func(exist bool)
+	// Promote promotes the 'cache node'.
+	Promote(n *Node)
 
-// PurgeFin will be called when corresponding cache object are released.
-// PurgeFin will be called after SetFin. If PurgeFin present DelFin will
-// not be executed but passed to the PurgeFin, it is up to the caller
-// to call it or not.
-type PurgeFin func(ns, key uint64, delfin DelFin)
+	// Ban evicts the 'cache node' and prevent subsequent 'promote'.
+	Ban(n *Node)
 
-// Cache is a cache tree.
-type Cache interface {
-	// SetCapacity sets cache capacity.
-	SetCapacity(capacity int)
+	// Evict evicts the 'cache node'.
+	Evict(n *Node)
 
-	// GetNamespace gets or creates a cache namespace for the given id.
-	GetNamespace(id uint64) Namespace
+	// EvictNS evicts 'cache node' with the given namespace.
+	EvictNS(ns uint64)
 
-	// Purge purges all cache namespaces, read Namespace.Purge method documentation.
-	Purge(fin PurgeFin)
+	// EvictAll evicts all 'cache node'.
+	EvictAll()
+
+	// Close closes the 'cache tree'
+	Close() error
+}
+
+// Value is a 'cacheable object'. It may implements util.Releaser, if
+// so the the Release method will be called once object is released.
+type Value interface{}
+
+type CacheGetter struct {
+	Cache *Cache
+	NS    uint64
+}
 
-	// Zap zaps all cache namespaces, read Namespace.Zap method documentation.
-	Zap(closed bool)
+func (g *CacheGetter) Get(key uint64, setFunc func() (size int, value Value)) *Handle {
+	return g.Cache.Get(g.NS, key, setFunc)
 }
 
-// Namespace is a cache namespace.
-type Namespace interface {
-	// Get gets cache object for the given key. The given SetFunc (if not nil) will
-	// be called if the given key does not exist.
-	// If the given key does not exist, SetFunc is nil or SetFunc return ok false, Get
-	// will return ok false.
-	Get(key uint64, setf SetFunc) (obj Object, ok bool)
+// The hash tables implementation is based on:
+// "Dynamic-Sized Nonblocking Hash Tables", by Yujie Liu, Kunlong Zhang, and Michael Spear. ACM Symposium on Principles of Distributed Computing, Jul 2014.
 
-	// Get deletes cache object for the given key. If exist the cache object will
-	// be deleted later when all of its handles have been released (i.e. no one use
-	// it anymore) and the given DelFin (if not nil) will finally be  executed. If
-	// such cache object does not exist the given DelFin will be executed anyway.
-	//
-	// Delete returns true if such cache object exist.
-	Delete(key uint64, fin DelFin) bool
+const (
+	mInitialSize           = 1 << 4
+	mOverflowThreshold     = 1 << 5
+	mOverflowGrowThreshold = 1 << 7
+)
 
-	// Purge deletes all cache objects, read Delete method documentation.
-	Purge(fin PurgeFin)
+type mBucket struct {
+	mu     sync.Mutex
+	node   []*Node
+	frozen bool
+}
 
-	// Zap detaches the namespace from the cache tree and delete all its cache
-	// objects. The cache objects deletion and finalizers execution are happen
-	// immediately, even if its existing handles haven't yet been released.
-	// A zapped namespace can't never be filled again.
-	// If closed is false then the Get function will always call the given SetFunc
-	// if it is not nil, but resultant of the SetFunc will not be cached.
-	Zap(closed bool)
+func (b *mBucket) freeze() []*Node {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	if !b.frozen {
+		b.frozen = true
+	}
+	return b.node
 }
 
-// Object is a cache object.
-type Object interface {
-	// Release releases the cache object. Other methods should not be called
-	// after the cache object has been released.
-	Release()
+func (b *mBucket) get(r *Cache, h *mNode, hash uint32, ns, key uint64, noset bool) (done, added bool, n *Node) {
+	b.mu.Lock()
+
+	if b.frozen {
+		b.mu.Unlock()
+		return
+	}
+
+	// Scan the node.
+	for _, n := range b.node {
+		if n.hash == hash && n.ns == ns && n.key == key {
+			atomic.AddInt32(&n.ref, 1)
+			b.mu.Unlock()
+			return true, false, n
+		}
+	}
+
+	// Get only.
+	if noset {
+		b.mu.Unlock()
+		return true, false, nil
+	}
+
+	// Create node.
+	n = &Node{
+		r:    r,
+		hash: hash,
+		ns:   ns,
+		key:  key,
+		ref:  1,
+	}
+	// Add node to bucket.
+	b.node = append(b.node, n)
+	bLen := len(b.node)
+	b.mu.Unlock()
+
+	// Update counter.
+	grow := atomic.AddInt32(&r.nodes, 1) >= h.growThreshold
+	if bLen > mOverflowThreshold {
+		grow = grow || atomic.AddInt32(&h.overflow, 1) >= mOverflowGrowThreshold
+	}
 
-	// Value returns value of the cache object.
-	Value() interface{}
+	// Grow.
+	if grow && atomic.CompareAndSwapInt32(&h.resizeInProgess, 0, 1) {
+		nhLen := len(h.buckets) << 1
+		nh := &mNode{
+			buckets:         make([]unsafe.Pointer, nhLen),
+			mask:            uint32(nhLen) - 1,
+			pred:            unsafe.Pointer(h),
+			growThreshold:   int32(nhLen * mOverflowThreshold),
+			shrinkThreshold: int32(nhLen >> 1),
+		}
+		ok := atomic.CompareAndSwapPointer(&r.mHead, unsafe.Pointer(h), unsafe.Pointer(nh))
+		if !ok {
+			panic("BUG: failed swapping head")
+		}
+		go nh.initBuckets()
+	}
+
+	return true, true, n
 }
 
-// Namespace state.
-type nsState int
+func (b *mBucket) delete(r *Cache, h *mNode, hash uint32, ns, key uint64) (done, deleted bool) {
+	b.mu.Lock()
 
-const (
-	nsEffective nsState = iota
-	nsZapped
-	nsClosed
-)
+	if b.frozen {
+		b.mu.Unlock()
+		return
+	}
 
-// Node state.
-type nodeState int
+	// Scan the node.
+	var (
+		n    *Node
+		bLen int
+	)
+	for i := range b.node {
+		n = b.node[i]
+		if n.ns == ns && n.key == key {
+			if atomic.LoadInt32(&n.ref) == 0 {
+				deleted = true
 
-const (
-	nodeEffective nodeState = iota
-	nodeEvicted
-	nodeRemoved
-)
+				// Call releaser.
+				if n.value != nil {
+					if r, ok := n.value.(util.Releaser); ok {
+						r.Release()
+					}
+					n.value = nil
+				}
+
+				// Remove node from bucket.
+				b.node = append(b.node[:i], b.node[i+1:]...)
+				bLen = len(b.node)
+			}
+			break
+		}
+	}
+	b.mu.Unlock()
 
-// Fake object.
-type fakeObject struct {
-	value interface{}
-	fin   func()
-	once  uint32
+	if deleted {
+		// Call OnDel.
+		for _, f := range n.onDel {
+			f()
+		}
+
+		// Update counter.
+		atomic.AddInt32(&r.size, int32(n.size)*-1)
+		shrink := atomic.AddInt32(&r.nodes, -1) < h.shrinkThreshold
+		if bLen >= mOverflowThreshold {
+			atomic.AddInt32(&h.overflow, -1)
+		}
+
+		// Shrink.
+		if shrink && len(h.buckets) > mInitialSize && atomic.CompareAndSwapInt32(&h.resizeInProgess, 0, 1) {
+			nhLen := len(h.buckets) >> 1
+			nh := &mNode{
+				buckets:         make([]unsafe.Pointer, nhLen),
+				mask:            uint32(nhLen) - 1,
+				pred:            unsafe.Pointer(h),
+				growThreshold:   int32(nhLen * mOverflowThreshold),
+				shrinkThreshold: int32(nhLen >> 1),
+			}
+			ok := atomic.CompareAndSwapPointer(&r.mHead, unsafe.Pointer(h), unsafe.Pointer(nh))
+			if !ok {
+				panic("BUG: failed swapping head")
+			}
+			go nh.initBuckets()
+		}
+	}
+
+	return true, deleted
 }
 
-func (o *fakeObject) Value() interface{} {
-	if atomic.LoadUint32(&o.once) == 0 {
-		return o.value
+type mNode struct {
+	buckets         []unsafe.Pointer // []*mBucket
+	mask            uint32
+	pred            unsafe.Pointer // *mNode
+	resizeInProgess int32
+
+	overflow        int32
+	growThreshold   int32
+	shrinkThreshold int32
+}
+
+func (n *mNode) initBucket(i uint32) *mBucket {
+	if b := (*mBucket)(atomic.LoadPointer(&n.buckets[i])); b != nil {
+		return b
+	}
+
+	p := (*mNode)(atomic.LoadPointer(&n.pred))
+	if p != nil {
+		var node []*Node
+		if n.mask > p.mask {
+			// Grow.
+			pb := (*mBucket)(atomic.LoadPointer(&p.buckets[i&p.mask]))
+			if pb == nil {
+				pb = p.initBucket(i & p.mask)
+			}
+			m := pb.freeze()
+			// Split nodes.
+			for _, x := range m {
+				if x.hash&n.mask == i {
+					node = append(node, x)
+				}
+			}
+		} else {
+			// Shrink.
+			pb0 := (*mBucket)(atomic.LoadPointer(&p.buckets[i]))
+			if pb0 == nil {
+				pb0 = p.initBucket(i)
+			}
+			pb1 := (*mBucket)(atomic.LoadPointer(&p.buckets[i+uint32(len(n.buckets))]))
+			if pb1 == nil {
+				pb1 = p.initBucket(i + uint32(len(n.buckets)))
+			}
+			m0 := pb0.freeze()
+			m1 := pb1.freeze()
+			// Merge nodes.
+			node = make([]*Node, 0, len(m0)+len(m1))
+			node = append(node, m0...)
+			node = append(node, m1...)
+		}
+		b := &mBucket{node: node}
+		if atomic.CompareAndSwapPointer(&n.buckets[i], nil, unsafe.Pointer(b)) {
+			if len(node) > mOverflowThreshold {
+				atomic.AddInt32(&n.overflow, int32(len(node)-mOverflowThreshold))
+			}
+			return b
+		}
+	}
+
+	return (*mBucket)(atomic.LoadPointer(&n.buckets[i]))
+}
+
+func (n *mNode) initBuckets() {
+	for i := range n.buckets {
+		n.initBucket(uint32(i))
+	}
+	atomic.StorePointer(&n.pred, nil)
+}
+
+// Cache is a 'cache map'.
+type Cache struct {
+	mu     sync.RWMutex
+	mHead  unsafe.Pointer // *mNode
+	nodes  int32
+	size   int32
+	cacher Cacher
+	closed bool
+}
+
+// NewCache creates a new 'cache map'. The cacher is optional and
+// may be nil.
+func NewCache(cacher Cacher) *Cache {
+	h := &mNode{
+		buckets:         make([]unsafe.Pointer, mInitialSize),
+		mask:            mInitialSize - 1,
+		growThreshold:   int32(mInitialSize * mOverflowThreshold),
+		shrinkThreshold: 0,
+	}
+	for i := range h.buckets {
+		h.buckets[i] = unsafe.Pointer(&mBucket{})
+	}
+	r := &Cache{
+		mHead:  unsafe.Pointer(h),
+		cacher: cacher,
+	}
+	return r
+}
+
+func (r *Cache) getBucket(hash uint32) (*mNode, *mBucket) {
+	h := (*mNode)(atomic.LoadPointer(&r.mHead))
+	i := hash & h.mask
+	b := (*mBucket)(atomic.LoadPointer(&h.buckets[i]))
+	if b == nil {
+		b = h.initBucket(i)
+	}
+	return h, b
+}
+
+func (r *Cache) delete(n *Node) bool {
+	for {
+		h, b := r.getBucket(n.hash)
+		done, deleted := b.delete(r, h, n.hash, n.ns, n.key)
+		if done {
+			return deleted
+		}
+	}
+	return false
+}
+
+// Nodes returns number of 'cache node' in the map.
+func (r *Cache) Nodes() int {
+	return int(atomic.LoadInt32(&r.nodes))
+}
+
+// Size returns sums of 'cache node' size in the map.
+func (r *Cache) Size() int {
+	return int(atomic.LoadInt32(&r.size))
+}
+
+// Capacity returns cache capacity.
+func (r *Cache) Capacity() int {
+	if r.cacher == nil {
+		return 0
+	}
+	return r.cacher.Capacity()
+}
+
+// SetCapacity sets cache capacity.
+func (r *Cache) SetCapacity(capacity int) {
+	if r.cacher != nil {
+		r.cacher.SetCapacity(capacity)
+	}
+}
+
+// Get gets 'cache node' with the given namespace and key.
+// If cache node is not found and setFunc is not nil, Get will atomically creates
+// the 'cache node' by calling setFunc. Otherwise Get will returns nil.
+//
+// The returned 'cache handle' should be released after use by calling Release
+// method.
+func (r *Cache) Get(ns, key uint64, setFunc func() (size int, value Value)) *Handle {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return nil
+	}
+
+	hash := murmur32(ns, key, 0xf00)
+	for {
+		h, b := r.getBucket(hash)
+		done, _, n := b.get(r, h, hash, ns, key, setFunc == nil)
+		if done {
+			if n != nil {
+				n.mu.Lock()
+				if n.value == nil {
+					if setFunc == nil {
+						n.mu.Unlock()
+						n.unref()
+						return nil
+					}
+
+					n.size, n.value = setFunc()
+					if n.value == nil {
+						n.size = 0
+						n.mu.Unlock()
+						n.unref()
+						return nil
+					}
+					atomic.AddInt32(&r.size, int32(n.size))
+				}
+				n.mu.Unlock()
+				if r.cacher != nil {
+					r.cacher.Promote(n)
+				}
+				return &Handle{unsafe.Pointer(n)}
+			}
+
+			break
+		}
 	}
 	return nil
 }
 
-func (o *fakeObject) Release() {
-	if !atomic.CompareAndSwapUint32(&o.once, 0, 1) {
+// Delete removes and ban 'cache node' with the given namespace and key.
+// A banned 'cache node' will never inserted into the 'cache tree'. Ban
+// only attributed to the particular 'cache node', so when a 'cache node'
+// is recreated it will not be banned.
+//
+// If onDel is not nil, then it will be executed if such 'cache node'
+// doesn't exist or once the 'cache node' is released.
+//
+// Delete return true is such 'cache node' exist.
+func (r *Cache) Delete(ns, key uint64, onDel func()) bool {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return false
+	}
+
+	hash := murmur32(ns, key, 0xf00)
+	for {
+		h, b := r.getBucket(hash)
+		done, _, n := b.get(r, h, hash, ns, key, true)
+		if done {
+			if n != nil {
+				if onDel != nil {
+					n.mu.Lock()
+					n.onDel = append(n.onDel, onDel)
+					n.mu.Unlock()
+				}
+				if r.cacher != nil {
+					r.cacher.Ban(n)
+				}
+				n.unref()
+				return true
+			}
+
+			break
+		}
+	}
+
+	if onDel != nil {
+		onDel()
+	}
+
+	return false
+}
+
+// Evict evicts 'cache node' with the given namespace and key. This will
+// simply call Cacher.Evict.
+//
+// Evict return true is such 'cache node' exist.
+func (r *Cache) Evict(ns, key uint64) bool {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return false
+	}
+
+	hash := murmur32(ns, key, 0xf00)
+	for {
+		h, b := r.getBucket(hash)
+		done, _, n := b.get(r, h, hash, ns, key, true)
+		if done {
+			if n != nil {
+				if r.cacher != nil {
+					r.cacher.Evict(n)
+				}
+				n.unref()
+				return true
+			}
+
+			break
+		}
+	}
+
+	return false
+}
+
+// EvictNS evicts 'cache node' with the given namespace. This will
+// simply call Cacher.EvictNS.
+func (r *Cache) EvictNS(ns uint64) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return
+	}
+
+	if r.cacher != nil {
+		r.cacher.EvictNS(ns)
+	}
+}
+
+// EvictAll evicts all 'cache node'. This will simply call Cacher.EvictAll.
+func (r *Cache) EvictAll() {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
 		return
 	}
-	if o.fin != nil {
-		o.fin()
-		o.fin = nil
+
+	if r.cacher != nil {
+		r.cacher.EvictAll()
+	}
+}
+
+// Close closes the 'cache map' and releases all 'cache node'.
+func (r *Cache) Close() error {
+	r.mu.Lock()
+	if !r.closed {
+		r.closed = true
+
+		if r.cacher != nil {
+			if err := r.cacher.Close(); err != nil {
+				return err
+			}
+		}
+
+		h := (*mNode)(r.mHead)
+		h.initBuckets()
+
+		for i := range h.buckets {
+			b := (*mBucket)(h.buckets[i])
+			for _, n := range b.node {
+				// Call releaser.
+				if n.value != nil {
+					if r, ok := n.value.(util.Releaser); ok {
+						r.Release()
+					}
+					n.value = nil
+				}
+
+				// Call OnDel.
+				for _, f := range n.onDel {
+					f()
+				}
+			}
+		}
 	}
+	r.mu.Unlock()
+	return nil
+}
+
+// Node is a 'cache node'.
+type Node struct {
+	r *Cache
+
+	hash    uint32
+	ns, key uint64
+
+	mu    sync.Mutex
+	size  int
+	value Value
+
+	ref   int32
+	onDel []func()
+
+	CacheData unsafe.Pointer
+}
+
+// NS returns this 'cache node' namespace.
+func (n *Node) NS() uint64 {
+	return n.ns
+}
+
+// Key returns this 'cache node' key.
+func (n *Node) Key() uint64 {
+	return n.key
+}
+
+// Size returns this 'cache node' size.
+func (n *Node) Size() int {
+	return n.size
+}
+
+// Value returns this 'cache node' value.
+func (n *Node) Value() Value {
+	return n.value
+}
+
+// Ref returns this 'cache node' ref counter.
+func (n *Node) Ref() int32 {
+	return atomic.LoadInt32(&n.ref)
+}
+
+// GetHandle returns an handle for this 'cache node'.
+func (n *Node) GetHandle() *Handle {
+	if atomic.AddInt32(&n.ref, 1) <= 1 {
+		panic("BUG: Node.GetHandle on zero ref")
+	}
+	return &Handle{unsafe.Pointer(n)}
+}
+
+func (n *Node) unref() {
+	if atomic.AddInt32(&n.ref, -1) == 0 {
+		n.r.delete(n)
+	}
+}
+
+func (n *Node) unrefLocked() {
+	if atomic.AddInt32(&n.ref, -1) == 0 {
+		n.r.mu.RLock()
+		if !n.r.closed {
+			n.r.delete(n)
+		}
+		n.r.mu.RUnlock()
+	}
+}
+
+type Handle struct {
+	n unsafe.Pointer // *Node
+}
+
+func (h *Handle) Value() Value {
+	n := (*Node)(atomic.LoadPointer(&h.n))
+	if n != nil {
+		return n.value
+	}
+	return nil
+}
+
+func (h *Handle) Release() {
+	nPtr := atomic.LoadPointer(&h.n)
+	if nPtr != nil && atomic.CompareAndSwapPointer(&h.n, nPtr, nil) {
+		n := (*Node)(nPtr)
+		n.unrefLocked()
+	}
+}
+
+func murmur32(ns, key uint64, seed uint32) uint32 {
+	const (
+		m = uint32(0x5bd1e995)
+		r = 24
+	)
+
+	k1 := uint32(ns >> 32)
+	k2 := uint32(ns)
+	k3 := uint32(key >> 32)
+	k4 := uint32(key)
+
+	k1 *= m
+	k1 ^= k1 >> r
+	k1 *= m
+
+	k2 *= m
+	k2 ^= k2 >> r
+	k2 *= m
+
+	k3 *= m
+	k3 ^= k3 >> r
+	k3 *= m
+
+	k4 *= m
+	k4 ^= k4 >> r
+	k4 *= m
+
+	h := seed
+
+	h *= m
+	h ^= k1
+	h *= m
+	h ^= k2
+	h *= m
+	h ^= k3
+	h *= m
+	h ^= k4
+
+	h ^= h >> 13
+	h *= m
+	h ^= h >> 15
+
+	return h
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache_test.go
index 07a9939b2..c2a50156f 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/cache_test.go
@@ -8,17 +8,289 @@ package cache
 
 import (
 	"math/rand"
+	"runtime"
+	"sync"
+	"sync/atomic"
 	"testing"
+	"time"
+	"unsafe"
 )
 
-func set(ns Namespace, key uint64, value interface{}, charge int, fin func()) Object {
-	obj, _ := ns.Get(key, func() (bool, interface{}, int, SetFin) {
-		return true, value, charge, fin
+type int32o int32
+
+func (o *int32o) acquire() {
+	if atomic.AddInt32((*int32)(o), 1) != 1 {
+		panic("BUG: invalid ref")
+	}
+}
+
+func (o *int32o) Release() {
+	if atomic.AddInt32((*int32)(o), -1) != 0 {
+		panic("BUG: invalid ref")
+	}
+}
+
+type releaserFunc struct {
+	fn    func()
+	value Value
+}
+
+func (r releaserFunc) Release() {
+	if r.fn != nil {
+		r.fn()
+	}
+}
+
+func set(c *Cache, ns, key uint64, value Value, charge int, relf func()) *Handle {
+	return c.Get(ns, key, func() (int, Value) {
+		if relf != nil {
+			return charge, releaserFunc{relf, value}
+		} else {
+			return charge, value
+		}
+	})
+}
+
+func TestCacheMap(t *testing.T) {
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	nsx := []struct {
+		nobjects, nhandles, concurrent, repeat int
+	}{
+		{10000, 400, 50, 3},
+		{100000, 1000, 100, 10},
+	}
+
+	var (
+		objects [][]int32o
+		handles [][]unsafe.Pointer
+	)
+
+	for _, x := range nsx {
+		objects = append(objects, make([]int32o, x.nobjects))
+		handles = append(handles, make([]unsafe.Pointer, x.nhandles))
+	}
+
+	c := NewCache(nil)
+
+	wg := new(sync.WaitGroup)
+	var done int32
+
+	for ns, x := range nsx {
+		for i := 0; i < x.concurrent; i++ {
+			wg.Add(1)
+			go func(ns, i, repeat int, objects []int32o, handles []unsafe.Pointer) {
+				defer wg.Done()
+				r := rand.New(rand.NewSource(time.Now().UnixNano()))
+
+				for j := len(objects) * repeat; j >= 0; j-- {
+					key := uint64(r.Intn(len(objects)))
+					h := c.Get(uint64(ns), key, func() (int, Value) {
+						o := &objects[key]
+						o.acquire()
+						return 1, o
+					})
+					if v := h.Value().(*int32o); v != &objects[key] {
+						t.Fatalf("#%d invalid value: want=%p got=%p", ns, &objects[key], v)
+					}
+					if objects[key] != 1 {
+						t.Fatalf("#%d invalid object %d: %d", ns, key, objects[key])
+					}
+					if !atomic.CompareAndSwapPointer(&handles[r.Intn(len(handles))], nil, unsafe.Pointer(h)) {
+						h.Release()
+					}
+				}
+			}(ns, i, x.repeat, objects[ns], handles[ns])
+		}
+
+		go func(handles []unsafe.Pointer) {
+			r := rand.New(rand.NewSource(time.Now().UnixNano()))
+
+			for atomic.LoadInt32(&done) == 0 {
+				i := r.Intn(len(handles))
+				h := (*Handle)(atomic.LoadPointer(&handles[i]))
+				if h != nil && atomic.CompareAndSwapPointer(&handles[i], unsafe.Pointer(h), nil) {
+					h.Release()
+				}
+				time.Sleep(time.Millisecond)
+			}
+		}(handles[ns])
+	}
+
+	go func() {
+		handles := make([]*Handle, 100000)
+		for atomic.LoadInt32(&done) == 0 {
+			for i := range handles {
+				handles[i] = c.Get(999999999, uint64(i), func() (int, Value) {
+					return 1, 1
+				})
+			}
+			for _, h := range handles {
+				h.Release()
+			}
+		}
+	}()
+
+	wg.Wait()
+
+	atomic.StoreInt32(&done, 1)
+
+	for _, handles0 := range handles {
+		for i := range handles0 {
+			h := (*Handle)(atomic.LoadPointer(&handles0[i]))
+			if h != nil && atomic.CompareAndSwapPointer(&handles0[i], unsafe.Pointer(h), nil) {
+				h.Release()
+			}
+		}
+	}
+
+	for ns, objects0 := range objects {
+		for i, o := range objects0 {
+			if o != 0 {
+				t.Fatalf("invalid object #%d.%d: ref=%d", ns, i, o)
+			}
+		}
+	}
+}
+
+func TestCacheMap_NodesAndSize(t *testing.T) {
+	c := NewCache(nil)
+	if c.Nodes() != 0 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 0, c.Nodes())
+	}
+	if c.Size() != 0 {
+		t.Errorf("invalid size counter: want=%d got=%d", 0, c.Size())
+	}
+	set(c, 0, 1, 1, 1, nil)
+	set(c, 0, 2, 2, 2, nil)
+	set(c, 1, 1, 3, 3, nil)
+	set(c, 2, 1, 4, 1, nil)
+	if c.Nodes() != 4 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 4, c.Nodes())
+	}
+	if c.Size() != 7 {
+		t.Errorf("invalid size counter: want=%d got=%d", 4, c.Size())
+	}
+}
+
+func TestLRUCache_Capacity(t *testing.T) {
+	c := NewCache(NewLRU(10))
+	if c.Capacity() != 10 {
+		t.Errorf("invalid capacity: want=%d got=%d", 10, c.Capacity())
+	}
+	set(c, 0, 1, 1, 1, nil).Release()
+	set(c, 0, 2, 2, 2, nil).Release()
+	set(c, 1, 1, 3, 3, nil).Release()
+	set(c, 2, 1, 4, 1, nil).Release()
+	set(c, 2, 2, 5, 1, nil).Release()
+	set(c, 2, 3, 6, 1, nil).Release()
+	set(c, 2, 4, 7, 1, nil).Release()
+	set(c, 2, 5, 8, 1, nil).Release()
+	if c.Nodes() != 7 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 7, c.Nodes())
+	}
+	if c.Size() != 10 {
+		t.Errorf("invalid size counter: want=%d got=%d", 10, c.Size())
+	}
+	c.SetCapacity(9)
+	if c.Capacity() != 9 {
+		t.Errorf("invalid capacity: want=%d got=%d", 9, c.Capacity())
+	}
+	if c.Nodes() != 6 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 6, c.Nodes())
+	}
+	if c.Size() != 8 {
+		t.Errorf("invalid size counter: want=%d got=%d", 8, c.Size())
+	}
+}
+
+func TestCacheMap_NilValue(t *testing.T) {
+	c := NewCache(NewLRU(10))
+	h := c.Get(0, 0, func() (size int, value Value) {
+		return 1, nil
 	})
-	return obj
+	if h != nil {
+		t.Error("cache handle is non-nil")
+	}
+	if c.Nodes() != 0 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 0, c.Nodes())
+	}
+	if c.Size() != 0 {
+		t.Errorf("invalid size counter: want=%d got=%d", 0, c.Size())
+	}
 }
 
-func TestCache_HitMiss(t *testing.T) {
+func TestLRUCache_GetLatency(t *testing.T) {
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	const (
+		concurrentSet = 30
+		concurrentGet = 3
+		duration      = 3 * time.Second
+		delay         = 3 * time.Millisecond
+		maxkey        = 100000
+	)
+
+	var (
+		set, getHit, getAll        int32
+		getMaxLatency, getDuration int64
+	)
+
+	c := NewCache(NewLRU(5000))
+	wg := &sync.WaitGroup{}
+	until := time.Now().Add(duration)
+	for i := 0; i < concurrentSet; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			r := rand.New(rand.NewSource(time.Now().UnixNano()))
+			for time.Now().Before(until) {
+				c.Get(0, uint64(r.Intn(maxkey)), func() (int, Value) {
+					time.Sleep(delay)
+					atomic.AddInt32(&set, 1)
+					return 1, 1
+				}).Release()
+			}
+		}(i)
+	}
+	for i := 0; i < concurrentGet; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			r := rand.New(rand.NewSource(time.Now().UnixNano()))
+			for {
+				mark := time.Now()
+				if mark.Before(until) {
+					h := c.Get(0, uint64(r.Intn(maxkey)), nil)
+					latency := int64(time.Now().Sub(mark))
+					m := atomic.LoadInt64(&getMaxLatency)
+					if latency > m {
+						atomic.CompareAndSwapInt64(&getMaxLatency, m, latency)
+					}
+					atomic.AddInt64(&getDuration, latency)
+					if h != nil {
+						atomic.AddInt32(&getHit, 1)
+						h.Release()
+					}
+					atomic.AddInt32(&getAll, 1)
+				} else {
+					break
+				}
+			}
+		}(i)
+	}
+
+	wg.Wait()
+	getAvglatency := time.Duration(getDuration) / time.Duration(getAll)
+	t.Logf("set=%d getHit=%d getAll=%d getMaxLatency=%v getAvgLatency=%v",
+		set, getHit, getAll, time.Duration(getMaxLatency), getAvglatency)
+
+	if getAvglatency > delay/3 {
+		t.Errorf("get avg latency > %v: got=%v", delay/3, getAvglatency)
+	}
+}
+
+func TestLRUCache_HitMiss(t *testing.T) {
 	cases := []struct {
 		key   uint64
 		value string
@@ -36,36 +308,37 @@ func TestCache_HitMiss(t *testing.T) {
 	}
 
 	setfin := 0
-	c := NewLRUCache(1000)
-	ns := c.GetNamespace(0)
+	c := NewCache(NewLRU(1000))
 	for i, x := range cases {
-		set(ns, x.key, x.value, len(x.value), func() {
+		set(c, 0, x.key, x.value, len(x.value), func() {
 			setfin++
 		}).Release()
 		for j, y := range cases {
-			r, ok := ns.Get(y.key, nil)
+			h := c.Get(0, y.key, nil)
 			if j <= i {
 				// should hit
-				if !ok {
+				if h == nil {
 					t.Errorf("case '%d' iteration '%d' is miss", i, j)
-				} else if r.Value().(string) != y.value {
-					t.Errorf("case '%d' iteration '%d' has invalid value got '%s', want '%s'", i, j, r.Value().(string), y.value)
+				} else {
+					if x := h.Value().(releaserFunc).value.(string); x != y.value {
+						t.Errorf("case '%d' iteration '%d' has invalid value got '%s', want '%s'", i, j, x, y.value)
+					}
 				}
 			} else {
 				// should miss
-				if ok {
-					t.Errorf("case '%d' iteration '%d' is hit , value '%s'", i, j, r.Value().(string))
+				if h != nil {
+					t.Errorf("case '%d' iteration '%d' is hit , value '%s'", i, j, h.Value().(releaserFunc).value.(string))
 				}
 			}
-			if ok {
-				r.Release()
+			if h != nil {
+				h.Release()
 			}
 		}
 	}
 
 	for i, x := range cases {
 		finalizerOk := false
-		ns.Delete(x.key, func(exist bool) {
+		c.Delete(0, x.key, func() {
 			finalizerOk = true
 		})
 
@@ -74,22 +347,24 @@ func TestCache_HitMiss(t *testing.T) {
 		}
 
 		for j, y := range cases {
-			r, ok := ns.Get(y.key, nil)
+			h := c.Get(0, y.key, nil)
 			if j > i {
 				// should hit
-				if !ok {
+				if h == nil {
 					t.Errorf("case '%d' iteration '%d' is miss", i, j)
-				} else if r.Value().(string) != y.value {
-					t.Errorf("case '%d' iteration '%d' has invalid value got '%s', want '%s'", i, j, r.Value().(string), y.value)
+				} else {
+					if x := h.Value().(releaserFunc).value.(string); x != y.value {
+						t.Errorf("case '%d' iteration '%d' has invalid value got '%s', want '%s'", i, j, x, y.value)
+					}
 				}
 			} else {
 				// should miss
-				if ok {
-					t.Errorf("case '%d' iteration '%d' is hit, value '%s'", i, j, r.Value().(string))
+				if h != nil {
+					t.Errorf("case '%d' iteration '%d' is hit, value '%s'", i, j, h.Value().(releaserFunc).value.(string))
 				}
 			}
-			if ok {
-				r.Release()
+			if h != nil {
+				h.Release()
 			}
 		}
 	}
@@ -100,137 +375,180 @@ func TestCache_HitMiss(t *testing.T) {
 }
 
 func TestLRUCache_Eviction(t *testing.T) {
-	c := NewLRUCache(12)
-	ns := c.GetNamespace(0)
-	o1 := set(ns, 1, 1, 1, nil)
-	set(ns, 2, 2, 1, nil).Release()
-	set(ns, 3, 3, 1, nil).Release()
-	set(ns, 4, 4, 1, nil).Release()
-	set(ns, 5, 5, 1, nil).Release()
-	if r, ok := ns.Get(2, nil); ok { // 1,3,4,5,2
-		r.Release()
-	}
-	set(ns, 9, 9, 10, nil).Release() // 5,2,9
-
-	for _, x := range []uint64{9, 2, 5, 1} {
-		r, ok := ns.Get(x, nil)
-		if !ok {
-			t.Errorf("miss for key '%d'", x)
+	c := NewCache(NewLRU(12))
+	o1 := set(c, 0, 1, 1, 1, nil)
+	set(c, 0, 2, 2, 1, nil).Release()
+	set(c, 0, 3, 3, 1, nil).Release()
+	set(c, 0, 4, 4, 1, nil).Release()
+	set(c, 0, 5, 5, 1, nil).Release()
+	if h := c.Get(0, 2, nil); h != nil { // 1,3,4,5,2
+		h.Release()
+	}
+	set(c, 0, 9, 9, 10, nil).Release() // 5,2,9
+
+	for _, key := range []uint64{9, 2, 5, 1} {
+		h := c.Get(0, key, nil)
+		if h == nil {
+			t.Errorf("miss for key '%d'", key)
 		} else {
-			if r.Value().(int) != int(x) {
-				t.Errorf("invalid value for key '%d' want '%d', got '%d'", x, x, r.Value().(int))
+			if x := h.Value().(int); x != int(key) {
+				t.Errorf("invalid value for key '%d' want '%d', got '%d'", key, key, x)
 			}
-			r.Release()
+			h.Release()
 		}
 	}
 	o1.Release()
-	for _, x := range []uint64{1, 2, 5} {
-		r, ok := ns.Get(x, nil)
-		if !ok {
-			t.Errorf("miss for key '%d'", x)
+	for _, key := range []uint64{1, 2, 5} {
+		h := c.Get(0, key, nil)
+		if h == nil {
+			t.Errorf("miss for key '%d'", key)
 		} else {
-			if r.Value().(int) != int(x) {
-				t.Errorf("invalid value for key '%d' want '%d', got '%d'", x, x, r.Value().(int))
+			if x := h.Value().(int); x != int(key) {
+				t.Errorf("invalid value for key '%d' want '%d', got '%d'", key, key, x)
 			}
-			r.Release()
+			h.Release()
 		}
 	}
-	for _, x := range []uint64{3, 4, 9} {
-		r, ok := ns.Get(x, nil)
-		if ok {
-			t.Errorf("hit for key '%d'", x)
-			if r.Value().(int) != int(x) {
-				t.Errorf("invalid value for key '%d' want '%d', got '%d'", x, x, r.Value().(int))
+	for _, key := range []uint64{3, 4, 9} {
+		h := c.Get(0, key, nil)
+		if h != nil {
+			t.Errorf("hit for key '%d'", key)
+			if x := h.Value().(int); x != int(key) {
+				t.Errorf("invalid value for key '%d' want '%d', got '%d'", key, key, x)
 			}
-			r.Release()
+			h.Release()
 		}
 	}
 }
 
-func TestLRUCache_SetGet(t *testing.T) {
-	c := NewLRUCache(13)
-	ns := c.GetNamespace(0)
-	for i := 0; i < 200; i++ {
-		n := uint64(rand.Intn(99999) % 20)
-		set(ns, n, n, 1, nil).Release()
-		if p, ok := ns.Get(n, nil); ok {
-			if p.Value() == nil {
-				t.Errorf("key '%d' contains nil value", n)
+func TestLRUCache_Evict(t *testing.T) {
+	c := NewCache(NewLRU(6))
+	set(c, 0, 1, 1, 1, nil).Release()
+	set(c, 0, 2, 2, 1, nil).Release()
+	set(c, 1, 1, 4, 1, nil).Release()
+	set(c, 1, 2, 5, 1, nil).Release()
+	set(c, 2, 1, 6, 1, nil).Release()
+	set(c, 2, 2, 7, 1, nil).Release()
+
+	for ns := 0; ns < 3; ns++ {
+		for key := 1; key < 3; key++ {
+			if h := c.Get(uint64(ns), uint64(key), nil); h != nil {
+				h.Release()
 			} else {
-				got := p.Value().(uint64)
-				if got != n {
-					t.Errorf("invalid value for key '%d' want '%d', got '%d'", n, n, got)
-				}
+				t.Errorf("Cache.Get on #%d.%d return nil", ns, key)
 			}
-			p.Release()
-		} else {
-			t.Errorf("key '%d' doesn't exist", n)
 		}
 	}
-}
 
-func TestLRUCache_Purge(t *testing.T) {
-	c := NewLRUCache(3)
-	ns1 := c.GetNamespace(0)
-	o1 := set(ns1, 1, 1, 1, nil)
-	o2 := set(ns1, 2, 2, 1, nil)
-	ns1.Purge(nil)
-	set(ns1, 3, 3, 1, nil).Release()
-	for _, x := range []uint64{1, 2, 3} {
-		r, ok := ns1.Get(x, nil)
-		if !ok {
-			t.Errorf("miss for key '%d'", x)
-		} else {
-			if r.Value().(int) != int(x) {
-				t.Errorf("invalid value for key '%d' want '%d', got '%d'", x, x, r.Value().(int))
+	if ok := c.Evict(0, 1); !ok {
+		t.Error("first Cache.Evict on #0.1 return false")
+	}
+	if ok := c.Evict(0, 1); ok {
+		t.Error("second Cache.Evict on #0.1 return true")
+	}
+	if h := c.Get(0, 1, nil); h != nil {
+		t.Errorf("Cache.Get on #0.1 return non-nil: %v", h.Value())
+	}
+
+	c.EvictNS(1)
+	if h := c.Get(1, 1, nil); h != nil {
+		t.Errorf("Cache.Get on #1.1 return non-nil: %v", h.Value())
+	}
+	if h := c.Get(1, 2, nil); h != nil {
+		t.Errorf("Cache.Get on #1.2 return non-nil: %v", h.Value())
+	}
+
+	c.EvictAll()
+	for ns := 0; ns < 3; ns++ {
+		for key := 1; key < 3; key++ {
+			if h := c.Get(uint64(ns), uint64(key), nil); h != nil {
+				t.Errorf("Cache.Get on #%d.%d return non-nil: %v", ns, key, h.Value())
 			}
-			r.Release()
 		}
 	}
-	o1.Release()
-	o2.Release()
-	for _, x := range []uint64{1, 2} {
-		r, ok := ns1.Get(x, nil)
-		if ok {
-			t.Errorf("hit for key '%d'", x)
-			if r.Value().(int) != int(x) {
-				t.Errorf("invalid value for key '%d' want '%d', got '%d'", x, x, r.Value().(int))
-			}
-			r.Release()
+}
+
+func TestLRUCache_Delete(t *testing.T) {
+	delFuncCalled := 0
+	delFunc := func() {
+		delFuncCalled++
+	}
+
+	c := NewCache(NewLRU(2))
+	set(c, 0, 1, 1, 1, nil).Release()
+	set(c, 0, 2, 2, 1, nil).Release()
+
+	if ok := c.Delete(0, 1, delFunc); !ok {
+		t.Error("Cache.Delete on #1 return false")
+	}
+	if h := c.Get(0, 1, nil); h != nil {
+		t.Errorf("Cache.Get on #1 return non-nil: %v", h.Value())
+	}
+	if ok := c.Delete(0, 1, delFunc); ok {
+		t.Error("Cache.Delete on #1 return true")
+	}
+
+	h2 := c.Get(0, 2, nil)
+	if h2 == nil {
+		t.Error("Cache.Get on #2 return nil")
+	}
+	if ok := c.Delete(0, 2, delFunc); !ok {
+		t.Error("(1) Cache.Delete on #2 return false")
+	}
+	if ok := c.Delete(0, 2, delFunc); !ok {
+		t.Error("(2) Cache.Delete on #2 return false")
+	}
+
+	set(c, 0, 3, 3, 1, nil).Release()
+	set(c, 0, 4, 4, 1, nil).Release()
+	c.Get(0, 2, nil).Release()
+
+	for key := 2; key <= 4; key++ {
+		if h := c.Get(0, uint64(key), nil); h != nil {
+			h.Release()
+		} else {
+			t.Errorf("Cache.Get on #%d return nil", key)
 		}
 	}
-}
 
-func BenchmarkLRUCache_SetRelease(b *testing.B) {
-	capacity := b.N / 100
-	if capacity <= 0 {
-		capacity = 10
+	h2.Release()
+	if h := c.Get(0, 2, nil); h != nil {
+		t.Errorf("Cache.Get on #2 return non-nil: %v", h.Value())
 	}
-	c := NewLRUCache(capacity)
-	ns := c.GetNamespace(0)
-	b.ResetTimer()
-	for i := uint64(0); i < uint64(b.N); i++ {
-		set(ns, i, nil, 1, nil).Release()
+
+	if delFuncCalled != 4 {
+		t.Errorf("delFunc isn't called 4 times: got=%d", delFuncCalled)
 	}
 }
 
-func BenchmarkLRUCache_SetReleaseTwice(b *testing.B) {
-	capacity := b.N / 100
-	if capacity <= 0 {
-		capacity = 10
+func TestLRUCache_Close(t *testing.T) {
+	relFuncCalled := 0
+	relFunc := func() {
+		relFuncCalled++
+	}
+	delFuncCalled := 0
+	delFunc := func() {
+		delFuncCalled++
 	}
-	c := NewLRUCache(capacity)
-	ns := c.GetNamespace(0)
-	b.ResetTimer()
 
-	na := b.N / 2
-	nb := b.N - na
+	c := NewCache(NewLRU(2))
+	set(c, 0, 1, 1, 1, relFunc).Release()
+	set(c, 0, 2, 2, 1, relFunc).Release()
 
-	for i := uint64(0); i < uint64(na); i++ {
-		set(ns, i, nil, 1, nil).Release()
+	h3 := set(c, 0, 3, 3, 1, relFunc)
+	if h3 == nil {
+		t.Error("Cache.Get on #3 return nil")
 	}
+	if ok := c.Delete(0, 3, delFunc); !ok {
+		t.Error("Cache.Delete on #3 return false")
+	}
+
+	c.Close()
 
-	for i := uint64(0); i < uint64(nb); i++ {
-		set(ns, i, nil, 1, nil).Release()
+	if relFuncCalled != 3 {
+		t.Errorf("relFunc isn't called 3 times: got=%d", relFuncCalled)
+	}
+	if delFuncCalled != 1 {
+		t.Errorf("delFunc isn't called 1 times: got=%d", delFuncCalled)
 	}
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/empty_cache.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/empty_cache.go
deleted file mode 100644
index 1fbf81459..000000000
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/empty_cache.go
+++ /dev/null
@@ -1,246 +0,0 @@
-// Copyright (c) 2013, Suryandaru Triandana <syndtr@gmail.com>
-// All rights reserved.
-//
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-package cache
-
-import (
-	"sync"
-	"sync/atomic"
-)
-
-type emptyCache struct {
-	sync.Mutex
-	table map[uint64]*emptyNS
-}
-
-// NewEmptyCache creates a new initialized empty cache.
-func NewEmptyCache() Cache {
-	return &emptyCache{
-		table: make(map[uint64]*emptyNS),
-	}
-}
-
-func (c *emptyCache) GetNamespace(id uint64) Namespace {
-	c.Lock()
-	defer c.Unlock()
-
-	if ns, ok := c.table[id]; ok {
-		return ns
-	}
-
-	ns := &emptyNS{
-		cache: c,
-		id:    id,
-		table: make(map[uint64]*emptyNode),
-	}
-	c.table[id] = ns
-	return ns
-}
-
-func (c *emptyCache) Purge(fin PurgeFin) {
-	c.Lock()
-	for _, ns := range c.table {
-		ns.purgeNB(fin)
-	}
-	c.Unlock()
-}
-
-func (c *emptyCache) Zap(closed bool) {
-	c.Lock()
-	for _, ns := range c.table {
-		ns.zapNB(closed)
-	}
-	c.table = make(map[uint64]*emptyNS)
-	c.Unlock()
-}
-
-func (*emptyCache) SetCapacity(capacity int) {}
-
-type emptyNS struct {
-	cache *emptyCache
-	id    uint64
-	table map[uint64]*emptyNode
-	state nsState
-}
-
-func (ns *emptyNS) Get(key uint64, setf SetFunc) (o Object, ok bool) {
-	ns.cache.Lock()
-
-	switch ns.state {
-	case nsZapped:
-		ns.cache.Unlock()
-		if setf == nil {
-			return
-		}
-
-		var value interface{}
-		var fin func()
-		ok, value, _, fin = setf()
-		if ok {
-			o = &fakeObject{
-				value: value,
-				fin:   fin,
-			}
-		}
-		return
-	case nsClosed:
-		ns.cache.Unlock()
-		return
-	}
-
-	n, ok := ns.table[key]
-	if ok {
-		n.ref++
-	} else {
-		if setf == nil {
-			ns.cache.Unlock()
-			return
-		}
-
-		var value interface{}
-		var fin func()
-		ok, value, _, fin = setf()
-		if !ok {
-			ns.cache.Unlock()
-			return
-		}
-
-		n = &emptyNode{
-			ns:     ns,
-			key:    key,
-			value:  value,
-			setfin: fin,
-			ref:    1,
-		}
-		ns.table[key] = n
-	}
-
-	ns.cache.Unlock()
-	o = &emptyObject{node: n}
-	return
-}
-
-func (ns *emptyNS) Delete(key uint64, fin DelFin) bool {
-	ns.cache.Lock()
-
-	if ns.state != nsEffective {
-		ns.cache.Unlock()
-		if fin != nil {
-			fin(false)
-		}
-		return false
-	}
-
-	n, ok := ns.table[key]
-	if !ok {
-		ns.cache.Unlock()
-		if fin != nil {
-			fin(false)
-		}
-		return false
-	}
-	n.delfin = fin
-	ns.cache.Unlock()
-	return true
-}
-
-func (ns *emptyNS) purgeNB(fin PurgeFin) {
-	if ns.state != nsEffective {
-		return
-	}
-	for _, n := range ns.table {
-		n.purgefin = fin
-	}
-}
-
-func (ns *emptyNS) Purge(fin PurgeFin) {
-	ns.cache.Lock()
-	ns.purgeNB(fin)
-	ns.cache.Unlock()
-}
-
-func (ns *emptyNS) zapNB(closed bool) {
-	if ns.state != nsEffective {
-		return
-	}
-	for _, n := range ns.table {
-		n.execFin()
-	}
-	if closed {
-		ns.state = nsClosed
-	} else {
-		ns.state = nsZapped
-	}
-	ns.table = nil
-}
-
-func (ns *emptyNS) Zap(closed bool) {
-	ns.cache.Lock()
-	ns.zapNB(closed)
-	delete(ns.cache.table, ns.id)
-	ns.cache.Unlock()
-}
-
-type emptyNode struct {
-	ns       *emptyNS
-	key      uint64
-	value    interface{}
-	ref      int
-	setfin   SetFin
-	delfin   DelFin
-	purgefin PurgeFin
-}
-
-func (n *emptyNode) execFin() {
-	if n.setfin != nil {
-		n.setfin()
-		n.setfin = nil
-	}
-	if n.purgefin != nil {
-		n.purgefin(n.ns.id, n.key, n.delfin)
-		n.delfin = nil
-		n.purgefin = nil
-	} else if n.delfin != nil {
-		n.delfin(true)
-		n.delfin = nil
-	}
-}
-
-func (n *emptyNode) evict() {
-	n.ns.cache.Lock()
-	n.ref--
-	if n.ref == 0 {
-		if n.ns.state == nsEffective {
-			// Remove elem.
-			delete(n.ns.table, n.key)
-			// Execute finalizer.
-			n.execFin()
-		}
-	} else if n.ref < 0 {
-		panic("leveldb/cache: emptyNode: negative node reference")
-	}
-	n.ns.cache.Unlock()
-}
-
-type emptyObject struct {
-	node *emptyNode
-	once uint32
-}
-
-func (o *emptyObject) Value() interface{} {
-	if atomic.LoadUint32(&o.once) == 0 {
-		return o.node.value
-	}
-	return nil
-}
-
-func (o *emptyObject) Release() {
-	if !atomic.CompareAndSwapUint32(&o.once, 0, 1) {
-		return
-	}
-	o.node.evict()
-	o.node = nil
-}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/lru.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/lru.go
new file mode 100644
index 000000000..d9a84cde1
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/lru.go
@@ -0,0 +1,195 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package cache
+
+import (
+	"sync"
+	"unsafe"
+)
+
+type lruNode struct {
+	n   *Node
+	h   *Handle
+	ban bool
+
+	next, prev *lruNode
+}
+
+func (n *lruNode) insert(at *lruNode) {
+	x := at.next
+	at.next = n
+	n.prev = at
+	n.next = x
+	x.prev = n
+}
+
+func (n *lruNode) remove() {
+	if n.prev != nil {
+		n.prev.next = n.next
+		n.next.prev = n.prev
+		n.prev = nil
+		n.next = nil
+	} else {
+		panic("BUG: removing removed node")
+	}
+}
+
+type lru struct {
+	mu       sync.Mutex
+	capacity int
+	used     int
+	recent   lruNode
+}
+
+func (r *lru) reset() {
+	r.recent.next = &r.recent
+	r.recent.prev = &r.recent
+	r.used = 0
+}
+
+func (r *lru) Capacity() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.capacity
+}
+
+func (r *lru) SetCapacity(capacity int) {
+	var evicted []*lruNode
+
+	r.mu.Lock()
+	r.capacity = capacity
+	for r.used > r.capacity {
+		rn := r.recent.prev
+		if rn == nil {
+			panic("BUG: invalid LRU used or capacity counter")
+		}
+		rn.remove()
+		rn.n.CacheData = nil
+		r.used -= rn.n.Size()
+		evicted = append(evicted, rn)
+	}
+	r.mu.Unlock()
+
+	for _, rn := range evicted {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) Promote(n *Node) {
+	var evicted []*lruNode
+
+	r.mu.Lock()
+	if n.CacheData == nil {
+		if n.Size() <= r.capacity {
+			rn := &lruNode{n: n, h: n.GetHandle()}
+			rn.insert(&r.recent)
+			n.CacheData = unsafe.Pointer(rn)
+			r.used += n.Size()
+
+			for r.used > r.capacity {
+				rn := r.recent.prev
+				if rn == nil {
+					panic("BUG: invalid LRU used or capacity counter")
+				}
+				rn.remove()
+				rn.n.CacheData = nil
+				r.used -= rn.n.Size()
+				evicted = append(evicted, rn)
+			}
+		}
+	} else {
+		rn := (*lruNode)(n.CacheData)
+		if !rn.ban {
+			rn.remove()
+			rn.insert(&r.recent)
+		}
+	}
+	r.mu.Unlock()
+
+	for _, rn := range evicted {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) Ban(n *Node) {
+	r.mu.Lock()
+	if n.CacheData == nil {
+		n.CacheData = unsafe.Pointer(&lruNode{n: n, ban: true})
+	} else {
+		rn := (*lruNode)(n.CacheData)
+		if !rn.ban {
+			rn.remove()
+			rn.ban = true
+			r.used -= rn.n.Size()
+			r.mu.Unlock()
+
+			rn.h.Release()
+			rn.h = nil
+			return
+		}
+	}
+	r.mu.Unlock()
+}
+
+func (r *lru) Evict(n *Node) {
+	r.mu.Lock()
+	rn := (*lruNode)(n.CacheData)
+	if rn == nil || rn.ban {
+		r.mu.Unlock()
+		return
+	}
+	n.CacheData = nil
+	r.mu.Unlock()
+
+	rn.h.Release()
+}
+
+func (r *lru) EvictNS(ns uint64) {
+	var evicted []*lruNode
+
+	r.mu.Lock()
+	for e := r.recent.prev; e != &r.recent; {
+		rn := e
+		e = e.prev
+		if rn.n.NS() == ns {
+			rn.remove()
+			rn.n.CacheData = nil
+			r.used -= rn.n.Size()
+			evicted = append(evicted, rn)
+		}
+	}
+	r.mu.Unlock()
+
+	for _, rn := range evicted {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) EvictAll() {
+	r.mu.Lock()
+	back := r.recent.prev
+	for rn := back; rn != &r.recent; rn = rn.prev {
+		rn.n.CacheData = nil
+	}
+	r.reset()
+	r.mu.Unlock()
+
+	for rn := back; rn != &r.recent; rn = rn.prev {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) Close() error {
+	return nil
+}
+
+// NewLRU create a new LRU-cache.
+func NewLRU(capacity int) Cacher {
+	r := &lru{capacity: capacity}
+	r.reset()
+	return r
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/lru_cache.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/lru_cache.go
deleted file mode 100644
index 3c98e076b..000000000
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/cache/lru_cache.go
+++ /dev/null
@@ -1,354 +0,0 @@
-// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
-// All rights reserved.
-//
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-package cache
-
-import (
-	"sync"
-	"sync/atomic"
-)
-
-// lruCache represent a LRU cache state.
-type lruCache struct {
-	sync.Mutex
-
-	recent   lruNode
-	table    map[uint64]*lruNs
-	capacity int
-	size     int
-}
-
-// NewLRUCache creates a new initialized LRU cache with the given capacity.
-func NewLRUCache(capacity int) Cache {
-	c := &lruCache{
-		table:    make(map[uint64]*lruNs),
-		capacity: capacity,
-	}
-	c.recent.rNext = &c.recent
-	c.recent.rPrev = &c.recent
-	return c
-}
-
-// SetCapacity set cache capacity.
-func (c *lruCache) SetCapacity(capacity int) {
-	c.Lock()
-	c.capacity = capacity
-	c.evict()
-	c.Unlock()
-}
-
-// GetNamespace return namespace object for given id.
-func (c *lruCache) GetNamespace(id uint64) Namespace {
-	c.Lock()
-	defer c.Unlock()
-
-	if p, ok := c.table[id]; ok {
-		return p
-	}
-
-	p := &lruNs{
-		lru:   c,
-		id:    id,
-		table: make(map[uint64]*lruNode),
-	}
-	c.table[id] = p
-	return p
-}
-
-// Purge purge entire cache.
-func (c *lruCache) Purge(fin PurgeFin) {
-	c.Lock()
-	for _, ns := range c.table {
-		ns.purgeNB(fin)
-	}
-	c.Unlock()
-}
-
-func (c *lruCache) Zap(closed bool) {
-	c.Lock()
-	for _, ns := range c.table {
-		ns.zapNB(closed)
-	}
-	c.table = make(map[uint64]*lruNs)
-	c.Unlock()
-}
-
-func (c *lruCache) evict() {
-	top := &c.recent
-	for n := c.recent.rPrev; c.size > c.capacity && n != top; {
-		n.state = nodeEvicted
-		n.rRemove()
-		n.evictNB()
-		c.size -= n.charge
-		n = c.recent.rPrev
-	}
-}
-
-type lruNs struct {
-	lru   *lruCache
-	id    uint64
-	table map[uint64]*lruNode
-	state nsState
-}
-
-func (ns *lruNs) Get(key uint64, setf SetFunc) (o Object, ok bool) {
-	lru := ns.lru
-	lru.Lock()
-
-	switch ns.state {
-	case nsZapped:
-		lru.Unlock()
-		if setf == nil {
-			return
-		}
-
-		var value interface{}
-		var fin func()
-		ok, value, _, fin = setf()
-		if ok {
-			o = &fakeObject{
-				value: value,
-				fin:   fin,
-			}
-		}
-		return
-	case nsClosed:
-		lru.Unlock()
-		return
-	}
-
-	n, ok := ns.table[key]
-	if ok {
-		switch n.state {
-		case nodeEvicted:
-			// Insert to recent list.
-			n.state = nodeEffective
-			n.ref++
-			lru.size += n.charge
-			lru.evict()
-			fallthrough
-		case nodeEffective:
-			// Bump to front
-			n.rRemove()
-			n.rInsert(&lru.recent)
-		}
-		n.ref++
-	} else {
-		if setf == nil {
-			lru.Unlock()
-			return
-		}
-
-		var value interface{}
-		var charge int
-		var fin func()
-		ok, value, charge, fin = setf()
-		if !ok {
-			lru.Unlock()
-			return
-		}
-
-		n = &lruNode{
-			ns:     ns,
-			key:    key,
-			value:  value,
-			charge: charge,
-			setfin: fin,
-			ref:    2,
-		}
-		ns.table[key] = n
-		n.rInsert(&lru.recent)
-
-		lru.size += charge
-		lru.evict()
-	}
-
-	lru.Unlock()
-	o = &lruObject{node: n}
-	return
-}
-
-func (ns *lruNs) Delete(key uint64, fin DelFin) bool {
-	lru := ns.lru
-	lru.Lock()
-
-	if ns.state != nsEffective {
-		lru.Unlock()
-		if fin != nil {
-			fin(false)
-		}
-		return false
-	}
-
-	n, ok := ns.table[key]
-	if !ok {
-		lru.Unlock()
-		if fin != nil {
-			fin(false)
-		}
-		return false
-	}
-
-	n.delfin = fin
-	switch n.state {
-	case nodeRemoved:
-		lru.Unlock()
-		return false
-	case nodeEffective:
-		lru.size -= n.charge
-		n.rRemove()
-		n.evictNB()
-	}
-	n.state = nodeRemoved
-
-	lru.Unlock()
-	return true
-}
-
-func (ns *lruNs) purgeNB(fin PurgeFin) {
-	lru := ns.lru
-	if ns.state != nsEffective {
-		return
-	}
-
-	for _, n := range ns.table {
-		n.purgefin = fin
-		if n.state == nodeEffective {
-			lru.size -= n.charge
-			n.rRemove()
-			n.evictNB()
-		}
-		n.state = nodeRemoved
-	}
-}
-
-func (ns *lruNs) Purge(fin PurgeFin) {
-	ns.lru.Lock()
-	ns.purgeNB(fin)
-	ns.lru.Unlock()
-}
-
-func (ns *lruNs) zapNB(closed bool) {
-	lru := ns.lru
-	if ns.state != nsEffective {
-		return
-	}
-
-	if closed {
-		ns.state = nsClosed
-	} else {
-		ns.state = nsZapped
-	}
-	for _, n := range ns.table {
-		if n.state == nodeEffective {
-			lru.size -= n.charge
-			n.rRemove()
-		}
-		n.state = nodeRemoved
-		n.execFin()
-	}
-	ns.table = nil
-}
-
-func (ns *lruNs) Zap(closed bool) {
-	ns.lru.Lock()
-	ns.zapNB(closed)
-	delete(ns.lru.table, ns.id)
-	ns.lru.Unlock()
-}
-
-type lruNode struct {
-	ns *lruNs
-
-	rNext, rPrev *lruNode
-
-	key      uint64
-	value    interface{}
-	charge   int
-	ref      int
-	state    nodeState
-	setfin   SetFin
-	delfin   DelFin
-	purgefin PurgeFin
-}
-
-func (n *lruNode) rInsert(at *lruNode) {
-	x := at.rNext
-	at.rNext = n
-	n.rPrev = at
-	n.rNext = x
-	x.rPrev = n
-}
-
-func (n *lruNode) rRemove() bool {
-	// only remove if not already removed
-	if n.rPrev == nil {
-		return false
-	}
-
-	n.rPrev.rNext = n.rNext
-	n.rNext.rPrev = n.rPrev
-	n.rPrev = nil
-	n.rNext = nil
-
-	return true
-}
-
-func (n *lruNode) execFin() {
-	if n.setfin != nil {
-		n.setfin()
-		n.setfin = nil
-	}
-	if n.purgefin != nil {
-		n.purgefin(n.ns.id, n.key, n.delfin)
-		n.delfin = nil
-		n.purgefin = nil
-	} else if n.delfin != nil {
-		n.delfin(true)
-		n.delfin = nil
-	}
-}
-
-func (n *lruNode) evictNB() {
-	n.ref--
-	if n.ref == 0 {
-		if n.ns.state == nsEffective {
-			// remove elem
-			delete(n.ns.table, n.key)
-			// execute finalizer
-			n.execFin()
-		}
-	} else if n.ref < 0 {
-		panic("leveldb/cache: lruCache: negative node reference")
-	}
-}
-
-func (n *lruNode) evict() {
-	n.ns.lru.Lock()
-	n.evictNB()
-	n.ns.lru.Unlock()
-}
-
-type lruObject struct {
-	node *lruNode
-	once uint32
-}
-
-func (o *lruObject) Value() interface{} {
-	if atomic.LoadUint32(&o.once) == 0 {
-		return o.node.value
-	}
-	return nil
-}
-
-func (o *lruObject) Release() {
-	if !atomic.CompareAndSwapUint32(&o.once, 0, 1) {
-		return
-	}
-
-	o.node.evict()
-	o.node = nil
-}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/config.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/config.go
deleted file mode 100644
index 511058897..000000000
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/config.go
+++ /dev/null
@@ -1,40 +0,0 @@
-// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
-// All rights reserved.
-//
-// Use of this source code is governed by a BSD-style license that can be
-// found in the LICENSE file.
-
-package leveldb
-
-const (
-	kNumLevels = 7
-
-	// Level-0 compaction is started when we hit this many files.
-	kL0_CompactionTrigger float64 = 4
-
-	// Soft limit on number of level-0 files.  We slow down writes at this point.
-	kL0_SlowdownWritesTrigger = 8
-
-	// Maximum number of level-0 files.  We stop writes at this point.
-	kL0_StopWritesTrigger = 12
-
-	// Maximum level to which a new compacted memdb is pushed if it
-	// does not create overlap.  We try to push to level 2 to avoid the
-	// relatively expensive level 0=>1 compactions and to avoid some
-	// expensive manifest file operations.  We do not push all the way to
-	// the largest level since that can generate a lot of wasted disk
-	// space if the same key space is being repeatedly overwritten.
-	kMaxMemCompactLevel = 2
-
-	// Maximum size of a table.
-	kMaxTableSize = 2 * 1048576
-
-	// Maximum bytes of overlaps in grandparent (i.e., level+2) before we
-	// stop building a single file in a level->level+1 compaction.
-	kMaxGrandParentOverlapBytes = 10 * kMaxTableSize
-
-	// Maximum number of bytes in all compacted files.  We avoid expanding
-	// the lower level file set of a compaction if it would make the
-	// total compaction cover more than this many bytes.
-	kExpCompactionMaxBytes = 25 * kMaxTableSize
-)
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/corrupt_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/corrupt_test.go
index a036e0893..a351874ed 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/corrupt_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/corrupt_test.go
@@ -9,13 +9,12 @@ package leveldb
 import (
 	"bytes"
 	"fmt"
+	"github.com/syndtr/goleveldb/leveldb/filter"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/storage"
 	"io"
 	"math/rand"
 	"testing"
-
-	"github.com/syndtr/goleveldb/leveldb/cache"
-	"github.com/syndtr/goleveldb/leveldb/opt"
-	"github.com/syndtr/goleveldb/leveldb/storage"
 )
 
 const ctValSize = 1000
@@ -32,8 +31,8 @@ func newDbCorruptHarnessWopt(t *testing.T, o *opt.Options) *dbCorruptHarness {
 
 func newDbCorruptHarness(t *testing.T) *dbCorruptHarness {
 	return newDbCorruptHarnessWopt(t, &opt.Options{
-		BlockCache: cache.NewLRUCache(100),
-		Strict:     opt.StrictJournalChecksum,
+		BlockCacheCapacity: 100,
+		Strict:             opt.StrictJournalChecksum,
 	})
 }
 
@@ -96,21 +95,22 @@ func (h *dbCorruptHarness) deleteRand(n, max int, rnd *rand.Rand) {
 	}
 }
 
-func (h *dbCorruptHarness) corrupt(ft storage.FileType, offset, n int) {
+func (h *dbCorruptHarness) corrupt(ft storage.FileType, fi, offset, n int) {
 	p := &h.dbHarness
 	t := p.t
 
-	var file storage.File
 	ff, _ := p.stor.GetFiles(ft)
-	for _, f := range ff {
-		if file == nil || f.Num() > file.Num() {
-			file = f
-		}
+	sff := files(ff)
+	sff.sort()
+	if fi < 0 {
+		fi = len(sff) - 1
 	}
-	if file == nil {
-		t.Fatalf("no such file with type %q", ft)
+	if fi >= len(sff) {
+		t.Fatalf("no such file with type %q with index %d", ft, fi)
 	}
 
+	file := sff[fi]
+
 	r, err := file.Open()
 	if err != nil {
 		t.Fatal("cannot open file: ", err)
@@ -225,8 +225,8 @@ func TestCorruptDB_Journal(t *testing.T) {
 	h.build(100)
 	h.check(100, 100)
 	h.closeDB()
-	h.corrupt(storage.TypeJournal, 19, 1)
-	h.corrupt(storage.TypeJournal, 32*1024+1000, 1)
+	h.corrupt(storage.TypeJournal, -1, 19, 1)
+	h.corrupt(storage.TypeJournal, -1, 32*1024+1000, 1)
 
 	h.openDB()
 	h.check(36, 36)
@@ -242,7 +242,7 @@ func TestCorruptDB_Table(t *testing.T) {
 	h.compactRangeAt(0, "", "")
 	h.compactRangeAt(1, "", "")
 	h.closeDB()
-	h.corrupt(storage.TypeTable, 100, 1)
+	h.corrupt(storage.TypeTable, -1, 100, 1)
 
 	h.openDB()
 	h.check(99, 99)
@@ -256,7 +256,7 @@ func TestCorruptDB_TableIndex(t *testing.T) {
 	h.build(10000)
 	h.compactMem()
 	h.closeDB()
-	h.corrupt(storage.TypeTable, -2000, 500)
+	h.corrupt(storage.TypeTable, -1, -2000, 500)
 
 	h.openDB()
 	h.check(5000, 9999)
@@ -267,9 +267,9 @@ func TestCorruptDB_TableIndex(t *testing.T) {
 func TestCorruptDB_MissingManifest(t *testing.T) {
 	rnd := rand.New(rand.NewSource(0x0badda7a))
 	h := newDbCorruptHarnessWopt(t, &opt.Options{
-		BlockCache:  cache.NewLRUCache(100),
-		Strict:      opt.StrictJournalChecksum,
-		WriteBuffer: 1000 * 60,
+		BlockCacheCapacity: 100,
+		Strict:             opt.StrictJournalChecksum,
+		WriteBuffer:        1000 * 60,
 	})
 
 	h.build(1000)
@@ -355,7 +355,7 @@ func TestCorruptDB_CorruptedManifest(t *testing.T) {
 	h.compactMem()
 	h.compactRange("", "")
 	h.closeDB()
-	h.corrupt(storage.TypeManifest, 0, 1000)
+	h.corrupt(storage.TypeManifest, -1, 0, 1000)
 	h.openAssert(false)
 
 	h.recover()
@@ -370,7 +370,7 @@ func TestCorruptDB_CompactionInputError(t *testing.T) {
 	h.build(10)
 	h.compactMem()
 	h.closeDB()
-	h.corrupt(storage.TypeTable, 100, 1)
+	h.corrupt(storage.TypeTable, -1, 100, 1)
 
 	h.openDB()
 	h.check(9, 9)
@@ -387,7 +387,7 @@ func TestCorruptDB_UnrelatedKeys(t *testing.T) {
 	h.build(10)
 	h.compactMem()
 	h.closeDB()
-	h.corrupt(storage.TypeTable, 100, 1)
+	h.corrupt(storage.TypeTable, -1, 100, 1)
 
 	h.openDB()
 	h.put(string(tkey(1000)), string(tval(1000, ctValSize)))
@@ -470,3 +470,31 @@ func TestCorruptDB_MissingTableFiles(t *testing.T) {
 
 	h.close()
 }
+
+func TestCorruptDB_RecoverTable(t *testing.T) {
+	h := newDbCorruptHarnessWopt(t, &opt.Options{
+		WriteBuffer:         112 * opt.KiB,
+		CompactionTableSize: 90 * opt.KiB,
+		Filter:              filter.NewBloomFilter(10),
+	})
+
+	h.build(1000)
+	h.compactMem()
+	h.compactRangeAt(0, "", "")
+	h.compactRangeAt(1, "", "")
+	seq := h.db.seq
+	h.closeDB()
+	h.corrupt(storage.TypeTable, 0, 1000, 1)
+	h.corrupt(storage.TypeTable, 3, 10000, 1)
+	// Corrupted filter shouldn't affect recovery.
+	h.corrupt(storage.TypeTable, 3, 113888, 10)
+	h.corrupt(storage.TypeTable, -1, 20000, 1)
+
+	h.recover()
+	if h.db.seq != seq {
+		t.Errorf("invalid seq, want=%d got=%d", seq, h.db.seq)
+	}
+	h.check(985, 985)
+
+	h.close()
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db.go
index 8e975dc63..323353b2a 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db.go
@@ -7,15 +7,17 @@
 package leveldb
 
 import (
-	"errors"
+	"container/list"
 	"fmt"
 	"io"
 	"os"
 	"runtime"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/journal"
 	"github.com/syndtr/goleveldb/leveldb/memdb"
@@ -30,41 +32,46 @@ type DB struct {
 	// Need 64-bit alignment.
 	seq uint64
 
+	// Session.
 	s *session
 
-	// MemDB
+	// MemDB.
 	memMu             sync.RWMutex
-	mem               *memdb.DB
-	frozenMem         *memdb.DB
+	memPool           chan *memdb.DB
+	mem, frozenMem    *memDB
 	journal           *journal.Writer
 	journalWriter     storage.Writer
 	journalFile       storage.File
 	frozenJournalFile storage.File
 	frozenSeq         uint64
 
-	// Snapshot
+	// Snapshot.
 	snapsMu   sync.Mutex
-	snapsRoot snapshotElement
+	snapsList *list.List
 
-	// Write
+	// Stats.
+	aliveSnaps, aliveIters int32
+
+	// Write.
 	writeC       chan *Batch
 	writeMergedC chan bool
 	writeLockC   chan struct{}
 	writeAckC    chan error
+	writeDelay   time.Duration
+	writeDelayN  int
 	journalC     chan *Batch
 	journalAckC  chan error
 
-	// Compaction
-	tcompCmdC     chan cCmd
-	tcompPauseC   chan chan<- struct{}
-	tcompTriggerC chan struct{}
-	mcompCmdC     chan cCmd
-	mcompTriggerC chan struct{}
-	compErrC      chan error
-	compErrSetC   chan error
-	compStats     [kNumLevels]cStats
-
-	// Close
+	// Compaction.
+	tcompCmdC   chan cCmd
+	tcompPauseC chan chan<- struct{}
+	mcompCmdC   chan cCmd
+	compErrC    chan error
+	compPerErrC chan error
+	compErrSetC chan error
+	compStats   []cStats
+
+	// Close.
 	closeW sync.WaitGroup
 	closeC chan struct{}
 	closed uint32
@@ -77,7 +84,11 @@ func openDB(s *session) (*DB, error) {
 	db := &DB{
 		s: s,
 		// Initial sequence
-		seq: s.stSeq,
+		seq: s.stSeqNum,
+		// MemDB
+		memPool: make(chan *memdb.DB, 1),
+		// Snapshot
+		snapsList: list.New(),
 		// Write
 		writeC:       make(chan *Batch),
 		writeMergedC: make(chan bool),
@@ -86,17 +97,16 @@ func openDB(s *session) (*DB, error) {
 		journalC:     make(chan *Batch),
 		journalAckC:  make(chan error),
 		// Compaction
-		tcompCmdC:     make(chan cCmd),
-		tcompPauseC:   make(chan chan<- struct{}),
-		tcompTriggerC: make(chan struct{}, 1),
-		mcompCmdC:     make(chan cCmd),
-		mcompTriggerC: make(chan struct{}, 1),
-		compErrC:      make(chan error),
-		compErrSetC:   make(chan error),
+		tcompCmdC:   make(chan cCmd),
+		tcompPauseC: make(chan chan<- struct{}),
+		mcompCmdC:   make(chan cCmd),
+		compErrC:    make(chan error),
+		compPerErrC: make(chan error),
+		compErrSetC: make(chan error),
+		compStats:   make([]cStats, s.o.GetNumLevel()),
 		// Close
 		closeC: make(chan struct{}),
 	}
-	db.initSnapshot()
 
 	if err := db.recoverJournal(); err != nil {
 		return nil, err
@@ -112,8 +122,9 @@ func openDB(s *session) (*DB, error) {
 		return nil, err
 	}
 
-	// Don't include compaction error goroutine into wait group.
+	// Doesn't need to be included in the wait group.
 	go db.compactionError()
+	go db.mpoolDrain()
 
 	db.closeW.Add(3)
 	go db.tCompaction()
@@ -135,9 +146,10 @@ func openDB(s *session) (*DB, error) {
 // detected in the DB. Corrupted DB can be recovered with Recover
 // function.
 //
+// The returned DB instance is goroutine-safe.
 // The DB must be closed after use, by calling Close method.
-func Open(p storage.Storage, o *opt.Options) (db *DB, err error) {
-	s, err := newSession(p, o)
+func Open(stor storage.Storage, o *opt.Options) (db *DB, err error) {
+	s, err := newSession(stor, o)
 	if err != nil {
 		return
 	}
@@ -177,6 +189,7 @@ func Open(p storage.Storage, o *opt.Options) (db *DB, err error) {
 // detected in the DB. Corrupted DB can be recovered with Recover
 // function.
 //
+// The returned DB instance is goroutine-safe.
 // The DB must be closed after use, by calling Close method.
 func OpenFile(path string, o *opt.Options) (db *DB, err error) {
 	stor, err := storage.OpenFile(path)
@@ -197,9 +210,10 @@ func OpenFile(path string, o *opt.Options) (db *DB, err error) {
 // The DB must already exist or it will returns an error.
 // Also, Recover will ignore ErrorIfMissing and ErrorIfExist options.
 //
+// The returned DB instance is goroutine-safe.
 // The DB must be closed after use, by calling Close method.
-func Recover(p storage.Storage, o *opt.Options) (db *DB, err error) {
-	s, err := newSession(p, o)
+func Recover(stor storage.Storage, o *opt.Options) (db *DB, err error) {
+	s, err := newSession(stor, o)
 	if err != nil {
 		return
 	}
@@ -225,6 +239,7 @@ func Recover(p storage.Storage, o *opt.Options) (db *DB, err error) {
 // RecoverFile uses standard file-system backed storage implementation as desribed
 // in the leveldb/storage package.
 //
+// The returned DB instance is goroutine-safe.
 // The DB must be closed after use, by calling Close method.
 func RecoverFile(path string, o *opt.Options) (db *DB, err error) {
 	stor, err := storage.OpenFile(path)
@@ -241,16 +256,28 @@ func RecoverFile(path string, o *opt.Options) (db *DB, err error) {
 }
 
 func recoverTable(s *session, o *opt.Options) error {
-	ff0, err := s.getFiles(storage.TypeTable)
+	o = dupOptions(o)
+	// Mask StrictReader, lets StrictRecovery doing its job.
+	o.Strict &= ^opt.StrictReader
+
+	// Get all tables and sort it by file number.
+	tableFiles_, err := s.getFiles(storage.TypeTable)
 	if err != nil {
 		return err
 	}
-	ff1 := files(ff0)
-	ff1.sort()
+	tableFiles := files(tableFiles_)
+	tableFiles.sort()
 
-	var mSeq uint64
-	var good, corrupted int
-	rec := new(sessionRecord)
+	var (
+		maxSeq                                                            uint64
+		recoveredKey, goodKey, corruptedKey, corruptedBlock, droppedTable int
+
+		// We will drop corrupted table.
+		strict = o.GetStrict(opt.StrictRecovery)
+
+		rec   = &sessionRecord{numLevel: o.GetNumLevel()}
+		bpool = util.NewBufferPool(o.GetBlockSize() + 5)
+	)
 	buildTable := func(iter iterator.Iterator) (tmp storage.File, size int64, err error) {
 		tmp = s.newTemp()
 		writer, err := tmp.Create()
@@ -264,8 +291,9 @@ func recoverTable(s *session, o *opt.Options) error {
 				tmp = nil
 			}
 		}()
+
+		// Copy entries.
 		tw := table.NewWriter(writer, o)
-		// Copy records.
 		for iter.Next() {
 			key := iter.Key()
 			if validIkey(key) {
@@ -296,45 +324,73 @@ func recoverTable(s *session, o *opt.Options) error {
 		if err != nil {
 			return err
 		}
-		defer reader.Close()
+		var closed bool
+		defer func() {
+			if !closed {
+				reader.Close()
+			}
+		}()
+
 		// Get file size.
 		size, err := reader.Seek(0, 2)
 		if err != nil {
 			return err
 		}
-		var tSeq uint64
-		var tgood, tcorrupted, blockerr int
-		var min, max []byte
-		tr := table.NewReader(reader, size, nil, o)
+
+		var (
+			tSeq                                     uint64
+			tgoodKey, tcorruptedKey, tcorruptedBlock int
+			imin, imax                               []byte
+		)
+		tr, err := table.NewReader(reader, size, storage.NewFileInfo(file), nil, bpool, o)
+		if err != nil {
+			return err
+		}
 		iter := tr.NewIterator(nil, nil)
-		iter.(iterator.ErrorCallbackSetter).SetErrorCallback(func(err error) {
-			s.logf("table@recovery found error @%d %q", file.Num(), err)
-			blockerr++
-		})
+		if itererr, ok := iter.(iterator.ErrorCallbackSetter); ok {
+			itererr.SetErrorCallback(func(err error) {
+				if errors.IsCorrupted(err) {
+					s.logf("table@recovery block corruption @%d %q", file.Num(), err)
+					tcorruptedBlock++
+				}
+			})
+		}
+
 		// Scan the table.
 		for iter.Next() {
 			key := iter.Key()
-			_, seq, _, ok := parseIkey(key)
-			if !ok {
-				tcorrupted++
+			_, seq, _, kerr := parseIkey(key)
+			if kerr != nil {
+				tcorruptedKey++
 				continue
 			}
-			tgood++
+			tgoodKey++
 			if seq > tSeq {
 				tSeq = seq
 			}
-			if min == nil {
-				min = append([]byte{}, key...)
+			if imin == nil {
+				imin = append([]byte{}, key...)
 			}
-			max = append(max[:0], key...)
+			imax = append(imax[:0], key...)
 		}
 		if err := iter.Error(); err != nil {
 			iter.Release()
 			return err
 		}
 		iter.Release()
-		if tgood > 0 {
-			if tcorrupted > 0 || blockerr > 0 {
+
+		goodKey += tgoodKey
+		corruptedKey += tcorruptedKey
+		corruptedBlock += tcorruptedBlock
+
+		if strict && (tcorruptedKey > 0 || tcorruptedBlock > 0) {
+			droppedTable++
+			s.logf("table@recovery dropped @%d Gk·%d Ck·%d Cb·%d S·%d Q·%d", file.Num(), tgoodKey, tcorruptedKey, tcorruptedBlock, size, tSeq)
+			return nil
+		}
+
+		if tgoodKey > 0 {
+			if tcorruptedKey > 0 || tcorruptedBlock > 0 {
 				// Rebuild the table.
 				s.logf("table@recovery rebuilding @%d", file.Num())
 				iter := tr.NewIterator(nil, nil)
@@ -343,62 +399,77 @@ func recoverTable(s *session, o *opt.Options) error {
 				if err != nil {
 					return err
 				}
+				closed = true
 				reader.Close()
 				if err := file.Replace(tmp); err != nil {
 					return err
 				}
 				size = newSize
 			}
-			if tSeq > mSeq {
-				mSeq = tSeq
+			if tSeq > maxSeq {
+				maxSeq = tSeq
 			}
+			recoveredKey += tgoodKey
 			// Add table to level 0.
-			rec.addTable(0, file.Num(), uint64(size), min, max)
-			s.logf("table@recovery recovered @%d N·%d C·%d B·%d S·%d Q·%d", file.Num(), tgood, tcorrupted, blockerr, size, tSeq)
+			rec.addTable(0, file.Num(), uint64(size), imin, imax)
+			s.logf("table@recovery recovered @%d Gk·%d Ck·%d Cb·%d S·%d Q·%d", file.Num(), tgoodKey, tcorruptedKey, tcorruptedBlock, size, tSeq)
 		} else {
-			s.logf("table@recovery unrecoverable @%d C·%d B·%d S·%d", file.Num(), tcorrupted, blockerr, size)
+			droppedTable++
+			s.logf("table@recovery unrecoverable @%d Ck·%d Cb·%d S·%d", file.Num(), tcorruptedKey, tcorruptedBlock, size)
 		}
 
-		good += tgood
-		corrupted += tcorrupted
-
 		return nil
 	}
+
 	// Recover all tables.
-	if len(ff1) > 0 {
-		s.logf("table@recovery F·%d", len(ff1))
-		s.markFileNum(ff1[len(ff1)-1].Num())
-		for _, file := range ff1 {
+	if len(tableFiles) > 0 {
+		s.logf("table@recovery F·%d", len(tableFiles))
+
+		// Mark file number as used.
+		s.markFileNum(tableFiles[len(tableFiles)-1].Num())
+
+		for _, file := range tableFiles {
 			if err := recoverTable(file); err != nil {
 				return err
 			}
 		}
-		s.logf("table@recovery recovered F·%d N·%d C·%d Q·%d", len(ff1), good, corrupted, mSeq)
+
+		s.logf("table@recovery recovered F·%d N·%d Gk·%d Ck·%d Q·%d", len(tableFiles), recoveredKey, goodKey, corruptedKey, maxSeq)
 	}
+
 	// Set sequence number.
-	rec.setSeq(mSeq + 1)
+	rec.setSeqNum(maxSeq)
+
 	// Create new manifest.
 	if err := s.create(); err != nil {
 		return err
 	}
+
 	// Commit.
 	return s.commit(rec)
 }
 
-func (d *DB) recoverJournal() error {
-	s := d.s
-
-	ff0, err := s.getFiles(storage.TypeJournal)
+func (db *DB) recoverJournal() error {
+	// Get all tables and sort it by file number.
+	journalFiles_, err := db.s.getFiles(storage.TypeJournal)
 	if err != nil {
 		return err
 	}
-	ff1 := files(ff0)
-	ff1.sort()
-	ff2 := make([]storage.File, 0, len(ff1))
-	for _, file := range ff1 {
-		if file.Num() >= s.stJournalNum || file.Num() == s.stPrevJournalNum {
-			s.markFileNum(file.Num())
-			ff2 = append(ff2, file)
+	journalFiles := files(journalFiles_)
+	journalFiles.sort()
+
+	// Discard older journal.
+	prev := -1
+	for i, file := range journalFiles {
+		if file.Num() >= db.s.stJournalNum {
+			if prev >= 0 {
+				i--
+				journalFiles[i] = journalFiles[prev]
+			}
+			journalFiles = journalFiles[i:]
+			break
+		} else if file.Num() == db.s.stPrevJournalNum {
+			prev = i
 		}
 	}
 
@@ -406,38 +477,43 @@ func (d *DB) recoverJournal() error {
 	var of storage.File
 	var mem *memdb.DB
 	batch := new(Batch)
-	cm := newCMem(s)
+	cm := newCMem(db.s)
 	buf := new(util.Buffer)
 	// Options.
-	strict := s.o.GetStrict(opt.StrictJournal)
-	checksum := s.o.GetStrict(opt.StrictJournalChecksum)
-	writeBuffer := s.o.GetWriteBuffer()
+	strict := db.s.o.GetStrict(opt.StrictJournal)
+	checksum := db.s.o.GetStrict(opt.StrictJournalChecksum)
+	writeBuffer := db.s.o.GetWriteBuffer()
 	recoverJournal := func(file storage.File) error {
-		s.logf("journal@recovery recovering @%d", file.Num())
+		db.logf("journal@recovery recovering @%d", file.Num())
 		reader, err := file.Open()
 		if err != nil {
 			return err
 		}
 		defer reader.Close()
+
+		// Create/reset journal reader instance.
 		if jr == nil {
-			jr = journal.NewReader(reader, dropper{s, file}, strict, checksum)
+			jr = journal.NewReader(reader, dropper{db.s, file}, strict, checksum)
 		} else {
-			jr.Reset(reader, dropper{s, file}, strict, checksum)
+			jr.Reset(reader, dropper{db.s, file}, strict, checksum)
 		}
+
+		// Flush memdb and remove obsolete journal file.
 		if of != nil {
 			if mem.Len() > 0 {
 				if err := cm.flush(mem, 0); err != nil {
 					return err
 				}
 			}
-			if err := cm.commit(file.Num(), d.seq); err != nil {
+			if err := cm.commit(file.Num(), db.seq); err != nil {
 				return err
 			}
 			cm.reset()
 			of.Remove()
 			of = nil
 		}
-		// Reset memdb.
+
+		// Replay journal to memdb.
 		mem.Reset()
 		for {
 			r, err := jr.Next()
@@ -445,43 +521,58 @@ func (d *DB) recoverJournal() error {
 				if err == io.EOF {
 					break
 				}
-				return err
+				return errors.SetFile(err, file)
 			}
+
 			buf.Reset()
 			if _, err := buf.ReadFrom(r); err != nil {
-				if strict {
-					return err
+				if err == io.ErrUnexpectedEOF {
+					// This is error returned due to corruption, with strict == false.
+					continue
+				} else {
+					return errors.SetFile(err, file)
 				}
-				continue
 			}
-			if err := batch.decode(buf.Bytes()); err != nil {
-				return err
-			}
-			if err := batch.memReplay(mem); err != nil {
-				return err
+			if err := batch.memDecodeAndReplay(db.seq, buf.Bytes(), mem); err != nil {
+				if strict || !errors.IsCorrupted(err) {
+					return errors.SetFile(err, file)
+				} else {
+					db.s.logf("journal error: %v (skipped)", err)
+					// We won't apply sequence number as it might be corrupted.
+					continue
+				}
 			}
-			d.seq = batch.seq + uint64(batch.len())
+
+			// Save sequence number.
+			db.seq = batch.seq + uint64(batch.Len())
+
+			// Flush it if large enough.
 			if mem.Size() >= writeBuffer {
-				// Large enough, flush it.
 				if err := cm.flush(mem, 0); err != nil {
 					return err
 				}
-				// Reset memdb.
 				mem.Reset()
 			}
 		}
+
 		of = file
 		return nil
 	}
+
 	// Recover all journals.
-	if len(ff2) > 0 {
-		s.logf("journal@recovery F·%d", len(ff2))
-		mem = memdb.New(s.icmp, writeBuffer)
-		for _, file := range ff2 {
+	if len(journalFiles) > 0 {
+		db.logf("journal@recovery F·%d", len(journalFiles))
+
+		// Mark file number as used.
+		db.s.markFileNum(journalFiles[len(journalFiles)-1].Num())
+
+		mem = memdb.New(db.s.icmp, writeBuffer)
+		for _, file := range journalFiles {
 			if err := recoverJournal(file); err != nil {
 				return err
 			}
 		}
+
 		// Flush the last journal.
 		if mem.Len() > 0 {
 			if err := cm.flush(mem, 0); err != nil {
@@ -489,72 +580,140 @@ func (d *DB) recoverJournal() error {
 			}
 		}
 	}
+
 	// Create a new journal.
-	if _, err := d.newMem(0); err != nil {
+	if _, err := db.newMem(0); err != nil {
 		return err
 	}
+
 	// Commit.
-	if err := cm.commit(d.journalFile.Num(), d.seq); err != nil {
+	if err := cm.commit(db.journalFile.Num(), db.seq); err != nil {
 		// Close journal.
-		if d.journal != nil {
-			d.journal.Close()
-			d.journalWriter.Close()
+		if db.journal != nil {
+			db.journal.Close()
+			db.journalWriter.Close()
 		}
 		return err
 	}
-	// Remove the last journal.
+
+	// Remove the last obsolete journal file.
 	if of != nil {
 		of.Remove()
 	}
+
 	return nil
 }
 
-func (d *DB) get(key []byte, seq uint64, ro *opt.ReadOptions) (value []byte, err error) {
-	s := d.s
+func (db *DB) get(key []byte, seq uint64, ro *opt.ReadOptions) (value []byte, err error) {
+	ikey := newIkey(key, seq, ktSeek)
 
-	ikey := newIKey(key, seq, tSeek)
-
-	em, fm := d.getMems()
-	for _, m := range [...]*memdb.DB{em, fm} {
+	em, fm := db.getMems()
+	for _, m := range [...]*memDB{em, fm} {
 		if m == nil {
 			continue
 		}
-		mk, mv, me := m.Find(ikey)
+		defer m.decref()
+
+		mk, mv, me := m.mdb.Find(ikey)
 		if me == nil {
-			ukey, _, t, ok := parseIkey(mk)
-			if ok && s.icmp.uCompare(ukey, key) == 0 {
-				if t == tDel {
+			ukey, _, kt, kerr := parseIkey(mk)
+			if kerr != nil {
+				// Shouldn't have had happen.
+				panic(kerr)
+			}
+			if db.s.icmp.uCompare(ukey, key) == 0 {
+				if kt == ktDel {
 					return nil, ErrNotFound
 				}
-				return mv, nil
+				return append([]byte{}, mv...), nil
 			}
 		} else if me != ErrNotFound {
 			return nil, me
 		}
 	}
 
-	v := s.version()
-	value, cSched, err := v.get(ikey, ro)
+	v := db.s.version()
+	value, cSched, err := v.get(ikey, ro, false)
 	v.release()
 	if cSched {
 		// Trigger table compaction.
-		d.compTrigger(d.tcompTriggerC)
+		db.compSendTrigger(db.tcompCmdC)
+	}
+	return
+}
+
+func (db *DB) has(key []byte, seq uint64, ro *opt.ReadOptions) (ret bool, err error) {
+	ikey := newIkey(key, seq, ktSeek)
+
+	em, fm := db.getMems()
+	for _, m := range [...]*memDB{em, fm} {
+		if m == nil {
+			continue
+		}
+		defer m.decref()
+
+		mk, _, me := m.mdb.Find(ikey)
+		if me == nil {
+			ukey, _, kt, kerr := parseIkey(mk)
+			if kerr != nil {
+				// Shouldn't have had happen.
+				panic(kerr)
+			}
+			if db.s.icmp.uCompare(ukey, key) == 0 {
+				if kt == ktDel {
+					return false, nil
+				}
+				return true, nil
+			}
+		} else if me != ErrNotFound {
+			return false, me
+		}
+	}
+
+	v := db.s.version()
+	_, cSched, err := v.get(ikey, ro, true)
+	v.release()
+	if cSched {
+		// Trigger table compaction.
+		db.compSendTrigger(db.tcompCmdC)
+	}
+	if err == nil {
+		ret = true
+	} else if err == ErrNotFound {
+		err = nil
 	}
 	return
 }
 
 // Get gets the value for the given key. It returns ErrNotFound if the
-// DB does not contain the key.
+// DB does not contains the key.
 //
-// The caller should not modify the contents of the returned slice, but
-// it is safe to modify the contents of the argument after Get returns.
-func (d *DB) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
-	err = d.ok()
+// The returned slice is its own copy, it is safe to modify the contents
+// of the returned slice.
+// It is safe to modify the contents of the argument after Get returns.
+func (db *DB) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
+	err = db.ok()
 	if err != nil {
 		return
 	}
 
-	return d.get(key, d.getSeq(), ro)
+	se := db.acquireSnapshot()
+	defer db.releaseSnapshot(se)
+	return db.get(key, se.seq, ro)
+}
+
+// Has returns true if the DB does contains the given key.
+//
+// It is safe to modify the contents of the argument after Get returns.
+func (db *DB) Has(key []byte, ro *opt.ReadOptions) (ret bool, err error) {
+	err = db.ok()
+	if err != nil {
+		return
+	}
+
+	se := db.acquireSnapshot()
+	defer db.releaseSnapshot(se)
+	return db.has(key, se.seq, ro)
 }
 
 // NewIterator returns an iterator for the latest snapshot of the
@@ -573,14 +732,16 @@ func (d *DB) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
 // The iterator must be released after use, by calling Release method.
 //
 // Also read Iterator documentation of the leveldb/iterator package.
-func (d *DB) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
-	if err := d.ok(); err != nil {
+func (db *DB) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
+	if err := db.ok(); err != nil {
 		return iterator.NewEmptyIterator(err)
 	}
 
-	p := d.newSnapshot()
-	defer p.Release()
-	return p.NewIterator(slice, ro)
+	se := db.acquireSnapshot()
+	defer db.releaseSnapshot(se)
+	// Iterator holds 'version' lock, 'version' is immutable so snapshot
+	// can be released after iterator created.
+	return db.newIterator(se.seq, slice, ro)
 }
 
 // GetSnapshot returns a latest snapshot of the underlying DB. A snapshot
@@ -588,25 +749,35 @@ func (d *DB) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterat
 // content of snapshot are guaranteed to be consistent.
 //
 // The snapshot must be released after use, by calling Release method.
-func (d *DB) GetSnapshot() (*Snapshot, error) {
-	if err := d.ok(); err != nil {
+func (db *DB) GetSnapshot() (*Snapshot, error) {
+	if err := db.ok(); err != nil {
 		return nil, err
 	}
 
-	return d.newSnapshot(), nil
+	return db.newSnapshot(), nil
 }
 
 // GetProperty returns value of the given property name.
 //
 // Property names:
 //	leveldb.num-files-at-level{n}
-//		Returns the number of filer at level 'n'.
+//		Returns the number of files at level 'n'.
 //	leveldb.stats
 //		Returns statistics of the underlying DB.
 //	leveldb.sstables
 //		Returns sstables list for each level.
-func (d *DB) GetProperty(name string) (value string, err error) {
-	err = d.ok()
+//	leveldb.blockpool
+//		Returns block pool stats.
+//	leveldb.cachedblock
+//		Returns size of cached block.
+//	leveldb.openedtables
+//		Returns number of opened tables.
+//	leveldb.alivesnaps
+//		Returns number of alive snapshots.
+//	leveldb.aliveiters
+//		Returns number of alive iterators.
+func (db *DB) GetProperty(name string) (value string, err error) {
+	err = db.ok()
 	if err != nil {
 		return
 	}
@@ -615,19 +786,18 @@ func (d *DB) GetProperty(name string) (value string, err error) {
 	if !strings.HasPrefix(name, prefix) {
 		return "", errors.New("leveldb: GetProperty: unknown property: " + name)
 	}
-
 	p := name[len(prefix):]
 
-	s := d.s
-	v := s.version()
+	v := db.s.version()
 	defer v.release()
 
+	numFilesPrefix := "num-files-at-level"
 	switch {
-	case strings.HasPrefix(p, "num-files-at-level"):
+	case strings.HasPrefix(p, numFilesPrefix):
 		var level uint
 		var rest string
-		n, _ := fmt.Scanf("%d%s", &level, &rest)
-		if n != 1 || level >= kNumLevels {
+		n, _ := fmt.Sscanf(p[len(numFilesPrefix):], "%d%s", &level, &rest)
+		if n != 1 || int(level) >= db.s.o.GetNumLevel() {
 			err = errors.New("leveldb: GetProperty: invalid property: " + name)
 		} else {
 			value = fmt.Sprint(v.tLen(int(level)))
@@ -636,22 +806,36 @@ func (d *DB) GetProperty(name string) (value string, err error) {
 		value = "Compactions\n" +
 			" Level |   Tables   |    Size(MB)   |    Time(sec)  |    Read(MB)   |   Write(MB)\n" +
 			"-------+------------+---------------+---------------+---------------+---------------\n"
-		for level, tt := range v.tables {
-			duration, read, write := d.compStats[level].get()
-			if len(tt) == 0 && duration == 0 {
+		for level, tables := range v.tables {
+			duration, read, write := db.compStats[level].get()
+			if len(tables) == 0 && duration == 0 {
 				continue
 			}
 			value += fmt.Sprintf(" %3d   | %10d | %13.5f | %13.5f | %13.5f | %13.5f\n",
-				level, len(tt), float64(tt.size())/1048576.0, duration.Seconds(),
+				level, len(tables), float64(tables.size())/1048576.0, duration.Seconds(),
 				float64(read)/1048576.0, float64(write)/1048576.0)
 		}
 	case p == "sstables":
-		for level, tt := range v.tables {
+		for level, tables := range v.tables {
 			value += fmt.Sprintf("--- level %d ---\n", level)
-			for _, t := range tt {
-				value += fmt.Sprintf("%d:%d[%q .. %q]\n", t.file.Num(), t.size, t.min, t.max)
+			for _, t := range tables {
+				value += fmt.Sprintf("%d:%d[%q .. %q]\n", t.file.Num(), t.size, t.imin, t.imax)
 			}
 		}
+	case p == "blockpool":
+		value = fmt.Sprintf("%v", db.s.tops.bpool)
+	case p == "cachedblock":
+		if db.s.tops.bcache != nil {
+			value = fmt.Sprintf("%d", db.s.tops.bcache.Size())
+		} else {
+			value = "<nil>"
+		}
+	case p == "openedtables":
+		value = fmt.Sprintf("%d", db.s.tops.cache.Size())
+	case p == "alivesnaps":
+		value = fmt.Sprintf("%d", atomic.LoadInt32(&db.aliveSnaps))
+	case p == "aliveiters":
+		value = fmt.Sprintf("%d", atomic.LoadInt32(&db.aliveIters))
 	default:
 		err = errors.New("leveldb: GetProperty: unknown property: " + name)
 	}
@@ -665,23 +849,23 @@ func (d *DB) GetProperty(name string) (value string, err error) {
 // data compresses by a factor of ten, the returned sizes will be one-tenth
 // the size of the corresponding user data size.
 // The results may not include the sizes of recently written data.
-func (d *DB) SizeOf(ranges []util.Range) (Sizes, error) {
-	if err := d.ok(); err != nil {
+func (db *DB) SizeOf(ranges []util.Range) (Sizes, error) {
+	if err := db.ok(); err != nil {
 		return nil, err
 	}
 
-	v := d.s.version()
+	v := db.s.version()
 	defer v.release()
 
 	sizes := make(Sizes, 0, len(ranges))
 	for _, r := range ranges {
-		min := newIKey(r.Start, kMaxSeq, tSeek)
-		max := newIKey(r.Limit, kMaxSeq, tSeek)
-		start, err := v.offsetOf(min)
+		imin := newIkey(r.Start, kMaxSeq, ktSeek)
+		imax := newIkey(r.Limit, kMaxSeq, ktSeek)
+		start, err := v.offsetOf(imin)
 		if err != nil {
 			return nil, err
 		}
-		limit, err := v.offsetOf(max)
+		limit, err := v.offsetOf(imax)
 		if err != nil {
 			return nil, err
 		}
@@ -695,61 +879,67 @@ func (d *DB) SizeOf(ranges []util.Range) (Sizes, error) {
 	return sizes, nil
 }
 
-// Close closes the DB. This will also releases any outstanding snapshot.
+// Close closes the DB. This will also releases any outstanding snapshot and
+// abort any in-flight compaction.
 //
 // It is not safe to close a DB until all outstanding iterators are released.
 // It is valid to call Close multiple times. Other methods should not be
 // called after the DB has been closed.
-func (d *DB) Close() error {
-	if !d.setClosed() {
+func (db *DB) Close() error {
+	if !db.setClosed() {
 		return ErrClosed
 	}
 
-	s := d.s
 	start := time.Now()
-	s.log("db@close closing")
+	db.log("db@close closing")
 
 	// Clear the finalizer.
-	runtime.SetFinalizer(d, nil)
+	runtime.SetFinalizer(db, nil)
 
 	// Get compaction error.
 	var err error
 	select {
-	case err = <-d.compErrC:
+	case err = <-db.compErrC:
 	default:
 	}
 
-	close(d.closeC)
+	// Signal all goroutines.
+	close(db.closeC)
 
-	// Wait for the close WaitGroup.
-	d.closeW.Wait()
+	// Wait for all gorotines to exit.
+	db.closeW.Wait()
+
+	// Lock writer and closes journal.
+	db.writeLockC <- struct{}{}
+	if db.journal != nil {
+		db.journal.Close()
+		db.journalWriter.Close()
+	}
 
-	// Close journal.
-	if d.journal != nil {
-		d.journal.Close()
-		d.journalWriter.Close()
+	if db.writeDelayN > 0 {
+		db.logf("db@write was delayed N·%d T·%v", db.writeDelayN, db.writeDelay)
 	}
 
 	// Close session.
-	s.close()
-	s.logf("db@close done T·%v", time.Since(start))
-	s.release()
+	db.s.close()
+	db.logf("db@close done T·%v", time.Since(start))
+	db.s.release()
 
-	if d.closer != nil {
-		if err1 := d.closer.Close(); err == nil {
+	if db.closer != nil {
+		if err1 := db.closer.Close(); err == nil {
 			err = err1
 		}
 	}
 
-	d.s = nil
-	d.mem = nil
-	d.frozenMem = nil
-	d.journal = nil
-	d.journalWriter = nil
-	d.journalFile = nil
-	d.frozenJournalFile = nil
-	d.snapsRoot = snapshotElement{}
-	d.closer = nil
+	// NIL'ing pointers.
+	db.s = nil
+	db.mem = nil
+	db.frozenMem = nil
+	db.journal = nil
+	db.journalWriter = nil
+	db.journalFile = nil
+	db.frozenJournalFile = nil
+	db.closer = nil
 
 	return err
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_compaction.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_compaction.go
index c82bd9f28..447407aba 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_compaction.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_compaction.go
@@ -7,11 +7,12 @@
 package leveldb
 
 import (
-	"errors"
 	"sync"
 	"time"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/memdb"
+	"github.com/syndtr/goleveldb/leveldb/opt"
 )
 
 var (
@@ -68,13 +69,13 @@ type cMem struct {
 }
 
 func newCMem(s *session) *cMem {
-	return &cMem{s: s, rec: new(sessionRecord)}
+	return &cMem{s: s, rec: &sessionRecord{numLevel: s.o.GetNumLevel()}}
 }
 
 func (c *cMem) flush(mem *memdb.DB, level int) error {
 	s := c.s
 
-	// Write memdb to table
+	// Write memdb to table.
 	iter := mem.NewIterator(nil)
 	defer iter.Release()
 	t, n, err := s.tops.createFrom(iter)
@@ -82,51 +83,85 @@ func (c *cMem) flush(mem *memdb.DB, level int) error {
 		return err
 	}
 
+	// Pick level.
 	if level < 0 {
-		level = s.version_NB().pickLevel(t.min.ukey(), t.max.ukey())
+		v := s.version()
+		level = v.pickLevel(t.imin.ukey(), t.imax.ukey())
+		v.release()
 	}
 	c.rec.addTableFile(level, t)
 
-	s.logf("mem@flush created L%d@%d N·%d S·%s %q:%q", level, t.file.Num(), n, shortenb(int(t.size)), t.min, t.max)
+	s.logf("mem@flush created L%d@%d N·%d S·%s %q:%q", level, t.file.Num(), n, shortenb(int(t.size)), t.imin, t.imax)
 
 	c.level = level
 	return nil
 }
 
 func (c *cMem) reset() {
-	c.rec = new(sessionRecord)
+	c.rec = &sessionRecord{numLevel: c.s.o.GetNumLevel()}
 }
 
 func (c *cMem) commit(journal, seq uint64) error {
 	c.rec.setJournalNum(journal)
-	c.rec.setSeq(seq)
-	// Commit changes
+	c.rec.setSeqNum(seq)
+
+	// Commit changes.
 	return c.s.commit(c.rec)
 }
 
-func (d *DB) compactionError() {
-	var err error
+func (db *DB) compactionError() {
+	var (
+		err     error
+		wlocked bool
+	)
 noerr:
+	// No error.
 	for {
 		select {
-		case _, _ = <-d.closeC:
-			return
-		case err = <-d.compErrSetC:
-			if err != nil {
+		case err = <-db.compErrSetC:
+			switch {
+			case err == nil:
+			case errors.IsCorrupted(err):
+				goto hasperr
+			default:
 				goto haserr
 			}
+		case _, _ = <-db.closeC:
+			return
 		}
 	}
 haserr:
+	// Transient error.
 	for {
 		select {
-		case _, _ = <-d.closeC:
-			return
-		case err = <-d.compErrSetC:
-			if err == nil {
+		case db.compErrC <- err:
+		case err = <-db.compErrSetC:
+			switch {
+			case err == nil:
 				goto noerr
+			case errors.IsCorrupted(err):
+				goto hasperr
+			default:
 			}
-		case d.compErrC <- err:
+		case _, _ = <-db.closeC:
+			return
+		}
+	}
+hasperr:
+	// Persistent error.
+	for {
+		select {
+		case db.compErrC <- err:
+		case db.compPerErrC <- err:
+		case db.writeLockC <- struct{}{}:
+			// Hold write lock, so that write won't pass-through.
+			wlocked = true
+		case _, _ = <-db.closeC:
+			if wlocked {
+				// We should release the lock or Close will hang.
+				<-db.writeLockC
+			}
+			return
 		}
 	}
 }
@@ -137,114 +172,159 @@ func (cnt *compactionTransactCounter) incr() {
 	*cnt++
 }
 
-func (d *DB) compactionTransact(name string, exec func(cnt *compactionTransactCounter) error, rollback func() error) {
-	s := d.s
+type compactionTransactInterface interface {
+	run(cnt *compactionTransactCounter) error
+	revert() error
+}
+
+func (db *DB) compactionTransact(name string, t compactionTransactInterface) {
 	defer func() {
 		if x := recover(); x != nil {
-			if x == errCompactionTransactExiting && rollback != nil {
-				if err := rollback(); err != nil {
-					s.logf("%s rollback error %q", name, err)
+			if x == errCompactionTransactExiting {
+				if err := t.revert(); err != nil {
+					db.logf("%s revert error %q", name, err)
 				}
 			}
 			panic(x)
 		}
 	}()
+
 	const (
 		backoffMin = 1 * time.Second
 		backoffMax = 8 * time.Second
 		backoffMul = 2 * time.Second
 	)
-	backoff := backoffMin
-	backoffT := time.NewTimer(backoff)
-	lastCnt := compactionTransactCounter(0)
+	var (
+		backoff  = backoffMin
+		backoffT = time.NewTimer(backoff)
+		lastCnt  = compactionTransactCounter(0)
+
+		disableBackoff = db.s.o.GetDisableCompactionBackoff()
+	)
 	for n := 0; ; n++ {
 		// Check wether the DB is closed.
-		if d.isClosed() {
-			s.logf("%s exiting", name)
-			d.compactionExitTransact()
+		if db.isClosed() {
+			db.logf("%s exiting", name)
+			db.compactionExitTransact()
 		} else if n > 0 {
-			s.logf("%s retrying N·%d", name, n)
+			db.logf("%s retrying N·%d", name, n)
 		}
 
 		// Execute.
 		cnt := compactionTransactCounter(0)
-		err := exec(&cnt)
+		err := t.run(&cnt)
+		if err != nil {
+			db.logf("%s error I·%d %q", name, cnt, err)
+		}
 
 		// Set compaction error status.
 		select {
-		case d.compErrSetC <- err:
-		case _, _ = <-d.closeC:
-			s.logf("%s exiting", name)
-			d.compactionExitTransact()
+		case db.compErrSetC <- err:
+		case perr := <-db.compPerErrC:
+			if err != nil {
+				db.logf("%s exiting (persistent error %q)", name, perr)
+				db.compactionExitTransact()
+			}
+		case _, _ = <-db.closeC:
+			db.logf("%s exiting", name)
+			db.compactionExitTransact()
 		}
 		if err == nil {
 			return
 		}
-		s.logf("%s error I·%d %q", name, cnt, err)
-
-		// Reset backoff duration if counter is advancing.
-		if cnt > lastCnt {
-			backoff = backoffMin
-			lastCnt = cnt
+		if errors.IsCorrupted(err) {
+			db.logf("%s exiting (corruption detected)", name)
+			db.compactionExitTransact()
 		}
 
-		// Backoff.
-		backoffT.Reset(backoff)
-		if backoff < backoffMax {
-			backoff *= backoffMul
-			if backoff > backoffMax {
-				backoff = backoffMax
+		if !disableBackoff {
+			// Reset backoff duration if counter is advancing.
+			if cnt > lastCnt {
+				backoff = backoffMin
+				lastCnt = cnt
+			}
+
+			// Backoff.
+			backoffT.Reset(backoff)
+			if backoff < backoffMax {
+				backoff *= backoffMul
+				if backoff > backoffMax {
+					backoff = backoffMax
+				}
+			}
+			select {
+			case <-backoffT.C:
+			case _, _ = <-db.closeC:
+				db.logf("%s exiting", name)
+				db.compactionExitTransact()
 			}
 		}
-		select {
-		case <-backoffT.C:
-		case _, _ = <-d.closeC:
-			s.logf("%s exiting", name)
-			d.compactionExitTransact()
-		}
 	}
 }
 
-func (d *DB) compactionExitTransact() {
+type compactionTransactFunc struct {
+	runFunc    func(cnt *compactionTransactCounter) error
+	revertFunc func() error
+}
+
+func (t *compactionTransactFunc) run(cnt *compactionTransactCounter) error {
+	return t.runFunc(cnt)
+}
+
+func (t *compactionTransactFunc) revert() error {
+	if t.revertFunc != nil {
+		return t.revertFunc()
+	}
+	return nil
+}
+
+func (db *DB) compactionTransactFunc(name string, run func(cnt *compactionTransactCounter) error, revert func() error) {
+	db.compactionTransact(name, &compactionTransactFunc{run, revert})
+}
+
+func (db *DB) compactionExitTransact() {
 	panic(errCompactionTransactExiting)
 }
 
-func (d *DB) memCompaction() {
-	mem := d.getFrozenMem()
+func (db *DB) memCompaction() {
+	mem := db.getFrozenMem()
 	if mem == nil {
 		return
 	}
+	defer mem.decref()
 
-	s := d.s
-	c := newCMem(s)
+	c := newCMem(db.s)
 	stats := new(cStatsStaging)
 
-	s.logf("mem@flush N·%d S·%s", mem.Len(), shortenb(mem.Size()))
+	db.logf("mem@flush N·%d S·%s", mem.mdb.Len(), shortenb(mem.mdb.Size()))
 
 	// Don't compact empty memdb.
-	if mem.Len() == 0 {
-		s.logf("mem@flush skipping")
+	if mem.mdb.Len() == 0 {
+		db.logf("mem@flush skipping")
 		// drop frozen mem
-		d.dropFrozenMem()
+		db.dropFrozenMem()
 		return
 	}
 
 	// Pause table compaction.
-	ch := make(chan struct{})
+	resumeC := make(chan struct{})
 	select {
-	case d.tcompPauseC <- (chan<- struct{})(ch):
-	case _, _ = <-d.closeC:
+	case db.tcompPauseC <- (chan<- struct{})(resumeC):
+	case <-db.compPerErrC:
+		close(resumeC)
+		resumeC = nil
+	case _, _ = <-db.closeC:
 		return
 	}
 
-	d.compactionTransact("mem@flush", func(cnt *compactionTransactCounter) (err error) {
+	db.compactionTransactFunc("mem@flush", func(cnt *compactionTransactCounter) (err error) {
 		stats.startTimer()
 		defer stats.stopTimer()
-		return c.flush(mem, -1)
+		return c.flush(mem.mdb, -1)
 	}, func() error {
 		for _, r := range c.rec.addedTables {
-			s.logf("mem@flush rollback @%d", r.num)
-			f := s.getTableFile(r.num)
+			db.logf("mem@flush revert @%d", r.num)
+			f := db.s.getTableFile(r.num)
 			if err := f.Remove(); err != nil {
 				return err
 			}
@@ -252,279 +332,327 @@ func (d *DB) memCompaction() {
 		return nil
 	})
 
-	d.compactionTransact("mem@commit", func(cnt *compactionTransactCounter) (err error) {
+	db.compactionTransactFunc("mem@commit", func(cnt *compactionTransactCounter) (err error) {
 		stats.startTimer()
 		defer stats.stopTimer()
-		return c.commit(d.journalFile.Num(), d.frozenSeq)
+		return c.commit(db.journalFile.Num(), db.frozenSeq)
 	}, nil)
 
-	s.logf("mem@flush commited F·%d T·%v", len(c.rec.addedTables), stats.duration)
+	db.logf("mem@flush committed F·%d T·%v", len(c.rec.addedTables), stats.duration)
 
 	for _, r := range c.rec.addedTables {
 		stats.write += r.size
 	}
-	d.compStats[c.level].add(stats)
+	db.compStats[c.level].add(stats)
 
 	// Drop frozen mem.
-	d.dropFrozenMem()
+	db.dropFrozenMem()
 
 	// Resume table compaction.
-	select {
-	case <-ch:
-	case _, _ = <-d.closeC:
-		return
+	if resumeC != nil {
+		select {
+		case <-resumeC:
+			close(resumeC)
+		case _, _ = <-db.closeC:
+			return
+		}
 	}
 
 	// Trigger table compaction.
-	d.compTrigger(d.mcompTriggerC)
+	db.compSendTrigger(db.tcompCmdC)
 }
 
-func (d *DB) tableCompaction(c *compaction, noTrivial bool) {
-	s := d.s
+type tableCompactionBuilder struct {
+	db           *DB
+	s            *session
+	c            *compaction
+	rec          *sessionRecord
+	stat0, stat1 *cStatsStaging
 
-	rec := new(sessionRecord)
-	rec.addCompactionPointer(c.level, c.max)
+	snapHasLastUkey bool
+	snapLastUkey    []byte
+	snapLastSeq     uint64
+	snapIter        int
+	snapKerrCnt     int
+	snapDropCnt     int
 
-	if !noTrivial && c.trivial() {
-		t := c.tables[0][0]
-		s.logf("table@move L%d@%d -> L%d", c.level, t.file.Num(), c.level+1)
-		rec.deleteTable(c.level, t.file.Num())
-		rec.addTableFile(c.level+1, t)
-		d.compactionTransact("table@move", func(cnt *compactionTransactCounter) (err error) {
-			return s.commit(rec)
-		}, nil)
-		return
-	}
+	kerrCnt int
+	dropCnt int
 
-	var stats [2]cStatsStaging
-	for i, tt := range c.tables {
-		for _, t := range tt {
-			stats[i].read += t.size
-			// Insert deleted tables into record
-			rec.deleteTable(c.level+i, t.file.Num())
-		}
-	}
-	sourceSize := int(stats[0].read + stats[1].read)
-	minSeq := d.minSeq()
-	s.logf("table@compaction L%d·%d -> L%d·%d S·%s Q·%d", c.level, len(c.tables[0]), c.level+1, len(c.tables[1]), shortenb(sourceSize), minSeq)
-
-	var snapUkey []byte
-	var snapHasUkey bool
-	var snapSeq uint64
-	var snapIter int
-	var snapDropCnt int
-	var dropCnt int
-	d.compactionTransact("table@build", func(cnt *compactionTransactCounter) (err error) {
-		ukey := append([]byte{}, snapUkey...)
-		hasUkey := snapHasUkey
-		lseq := snapSeq
-		dropCnt = snapDropCnt
-		snapSched := snapIter == 0
-
-		var tw *tWriter
-		finish := func() error {
-			t, err := tw.finish()
-			if err != nil {
-				return err
+	minSeq    uint64
+	strict    bool
+	tableSize int
+
+	tw *tWriter
+}
+
+func (b *tableCompactionBuilder) appendKV(key, value []byte) error {
+	// Create new table if not already.
+	if b.tw == nil {
+		// Check for pause event.
+		if b.db != nil {
+			select {
+			case ch := <-b.db.tcompPauseC:
+				b.db.pauseCompaction(ch)
+			case _, _ = <-b.db.closeC:
+				b.db.compactionExitTransact()
+			default:
 			}
-			rec.addTableFile(c.level+1, t)
-			stats[1].write += t.size
-			s.logf("table@build created L%d@%d N·%d S·%s %q:%q", c.level+1, t.file.Num(), tw.tw.EntriesLen(), shortenb(int(t.size)), t.min, t.max)
-			return nil
 		}
 
-		defer func() {
-			stats[1].stopTimer()
-			if tw != nil {
-				tw.drop()
-				tw = nil
-			}
-		}()
+		// Create new table.
+		var err error
+		b.tw, err = b.s.tops.create()
+		if err != nil {
+			return err
+		}
+	}
 
-		stats[1].startTimer()
-		iter := c.newIterator()
-		defer iter.Release()
-		for i := 0; iter.Next(); i++ {
-			// Incr transact counter.
-			cnt.incr()
-
-			// Skip until last state.
-			if i < snapIter {
-				continue
-			}
+	// Write key/value into table.
+	return b.tw.append(key, value)
+}
 
-			key := iKey(iter.Key())
+func (b *tableCompactionBuilder) needFlush() bool {
+	return b.tw.tw.BytesLen() >= b.tableSize
+}
 
-			if c.shouldStopBefore(key) && tw != nil {
-				err = finish()
-				if err != nil {
-					return
-				}
-				snapSched = true
-				tw = nil
-			}
+func (b *tableCompactionBuilder) flush() error {
+	t, err := b.tw.finish()
+	if err != nil {
+		return err
+	}
+	b.rec.addTableFile(b.c.level+1, t)
+	b.stat1.write += t.size
+	b.s.logf("table@build created L%d@%d N·%d S·%s %q:%q", b.c.level+1, t.file.Num(), b.tw.tw.EntriesLen(), shortenb(int(t.size)), t.imin, t.imax)
+	b.tw = nil
+	return nil
+}
 
-			// Scheduled for snapshot, snapshot will used to retry compaction
-			// if error occured.
-			if snapSched {
-				snapUkey = append(snapUkey[:0], ukey...)
-				snapHasUkey = hasUkey
-				snapSeq = lseq
-				snapIter = i
-				snapDropCnt = dropCnt
-				snapSched = false
-			}
+func (b *tableCompactionBuilder) cleanup() {
+	if b.tw != nil {
+		b.tw.drop()
+		b.tw = nil
+	}
+}
 
-			if seq, t, ok := key.parseNum(); !ok {
-				// Don't drop error keys
-				ukey = ukey[:0]
-				hasUkey = false
-				lseq = kMaxSeq
-			} else {
-				if !hasUkey || s.icmp.uCompare(key.ukey(), ukey) != 0 {
-					// First occurrence of this user key
-					ukey = append(ukey[:0], key.ukey()...)
-					hasUkey = true
-					lseq = kMaxSeq
-				}
+func (b *tableCompactionBuilder) run(cnt *compactionTransactCounter) error {
+	snapResumed := b.snapIter > 0
+	hasLastUkey := b.snapHasLastUkey // The key might has zero length, so this is necessary.
+	lastUkey := append([]byte{}, b.snapLastUkey...)
+	lastSeq := b.snapLastSeq
+	b.kerrCnt = b.snapKerrCnt
+	b.dropCnt = b.snapDropCnt
+	// Restore compaction state.
+	b.c.restore()
 
-				drop := false
-				if lseq <= minSeq {
-					// Dropped because newer entry for same user key exist
-					drop = true // (A)
-				} else if t == tDel && seq <= minSeq && c.isBaseLevelForKey(ukey) {
-					// For this user key:
-					// (1) there is no data in higher levels
-					// (2) data in lower levels will have larger seq numbers
-					// (3) data in layers that are being compacted here and have
-					//     smaller seq numbers will be dropped in the next
-					//     few iterations of this loop (by rule (A) above).
-					// Therefore this deletion marker is obsolete and can be dropped.
-					drop = true
-				}
+	defer b.cleanup()
 
-				lseq = seq
-				if drop {
-					dropCnt++
-					continue
-				}
-			}
+	b.stat1.startTimer()
+	defer b.stat1.stopTimer()
 
-			// Create new table if not already
-			if tw == nil {
-				// Check for pause event.
-				select {
-				case ch := <-d.tcompPauseC:
-					d.pauseCompaction(ch)
-				case _, _ = <-d.closeC:
-					d.compactionExitTransact()
-				default:
-				}
+	iter := b.c.newIterator()
+	defer iter.Release()
+	for i := 0; iter.Next(); i++ {
+		// Incr transact counter.
+		cnt.incr()
+
+		// Skip until last state.
+		if i < b.snapIter {
+			continue
+		}
 
-				// Create new table.
-				tw, err = s.tops.create()
-				if err != nil {
-					return
+		resumed := false
+		if snapResumed {
+			resumed = true
+			snapResumed = false
+		}
+
+		ikey := iter.Key()
+		ukey, seq, kt, kerr := parseIkey(ikey)
+
+		if kerr == nil {
+			shouldStop := !resumed && b.c.shouldStopBefore(ikey)
+
+			if !hasLastUkey || b.s.icmp.uCompare(lastUkey, ukey) != 0 {
+				// First occurrence of this user key.
+
+				// Only rotate tables if ukey doesn't hop across.
+				if b.tw != nil && (shouldStop || b.needFlush()) {
+					if err := b.flush(); err != nil {
+						return err
+					}
+
+					// Creates snapshot of the state.
+					b.c.save()
+					b.snapHasLastUkey = hasLastUkey
+					b.snapLastUkey = append(b.snapLastUkey[:0], lastUkey...)
+					b.snapLastSeq = lastSeq
+					b.snapIter = i
+					b.snapKerrCnt = b.kerrCnt
+					b.snapDropCnt = b.dropCnt
 				}
-			}
 
-			// Write key/value into table
-			err = tw.add(key, iter.Value())
-			if err != nil {
-				return
+				hasLastUkey = true
+				lastUkey = append(lastUkey[:0], ukey...)
+				lastSeq = kMaxSeq
 			}
 
-			// Finish table if it is big enough
-			if tw.tw.BytesLen() >= kMaxTableSize {
-				err = finish()
-				if err != nil {
-					return
-				}
-				snapSched = true
-				tw = nil
+			switch {
+			case lastSeq <= b.minSeq:
+				// Dropped because newer entry for same user key exist
+				fallthrough // (A)
+			case kt == ktDel && seq <= b.minSeq && b.c.baseLevelForKey(lastUkey):
+				// For this user key:
+				// (1) there is no data in higher levels
+				// (2) data in lower levels will have larger seq numbers
+				// (3) data in layers that are being compacted here and have
+				//     smaller seq numbers will be dropped in the next
+				//     few iterations of this loop (by rule (A) above).
+				// Therefore this deletion marker is obsolete and can be dropped.
+				lastSeq = seq
+				b.dropCnt++
+				continue
+			default:
+				lastSeq = seq
+			}
+		} else {
+			if b.strict {
+				return kerr
 			}
+
+			// Don't drop corrupted keys.
+			hasLastUkey = false
+			lastUkey = lastUkey[:0]
+			lastSeq = kMaxSeq
+			b.kerrCnt++
 		}
 
-		err = iter.Error()
-		if err != nil {
-			return
+		if err := b.appendKV(ikey, iter.Value()); err != nil {
+			return err
 		}
+	}
 
-		// Finish last table
-		if tw != nil && !tw.empty() {
-			err = finish()
-			if err != nil {
-				return
-			}
-			tw = nil
+	if err := iter.Error(); err != nil {
+		return err
+	}
+
+	// Finish last table.
+	if b.tw != nil && !b.tw.empty() {
+		return b.flush()
+	}
+	return nil
+}
+
+func (b *tableCompactionBuilder) revert() error {
+	for _, at := range b.rec.addedTables {
+		b.s.logf("table@build revert @%d", at.num)
+		f := b.s.getTableFile(at.num)
+		if err := f.Remove(); err != nil {
+			return err
 		}
+	}
+	return nil
+}
+
+func (db *DB) tableCompaction(c *compaction, noTrivial bool) {
+	defer c.release()
+
+	rec := &sessionRecord{numLevel: db.s.o.GetNumLevel()}
+	rec.addCompPtr(c.level, c.imax)
+
+	if !noTrivial && c.trivial() {
+		t := c.tables[0][0]
+		db.logf("table@move L%d@%d -> L%d", c.level, t.file.Num(), c.level+1)
+		rec.delTable(c.level, t.file.Num())
+		rec.addTableFile(c.level+1, t)
+		db.compactionTransactFunc("table@move", func(cnt *compactionTransactCounter) (err error) {
+			return db.s.commit(rec)
+		}, nil)
 		return
-	}, func() error {
-		for _, r := range rec.addedTables {
-			s.logf("table@build rollback @%d", r.num)
-			f := s.getTableFile(r.num)
-			if err := f.Remove(); err != nil {
-				return err
-			}
+	}
+
+	var stats [2]cStatsStaging
+	for i, tables := range c.tables {
+		for _, t := range tables {
+			stats[i].read += t.size
+			// Insert deleted tables into record
+			rec.delTable(c.level+i, t.file.Num())
 		}
-		return nil
-	})
+	}
+	sourceSize := int(stats[0].read + stats[1].read)
+	minSeq := db.minSeq()
+	db.logf("table@compaction L%d·%d -> L%d·%d S·%s Q·%d", c.level, len(c.tables[0]), c.level+1, len(c.tables[1]), shortenb(sourceSize), minSeq)
+
+	b := &tableCompactionBuilder{
+		db:        db,
+		s:         db.s,
+		c:         c,
+		rec:       rec,
+		stat1:     &stats[1],
+		minSeq:    minSeq,
+		strict:    db.s.o.GetStrict(opt.StrictCompaction),
+		tableSize: db.s.o.GetCompactionTableSize(c.level + 1),
+	}
+	db.compactionTransact("table@build", b)
 
 	// Commit changes
-	d.compactionTransact("table@commit", func(cnt *compactionTransactCounter) (err error) {
+	db.compactionTransactFunc("table@commit", func(cnt *compactionTransactCounter) (err error) {
 		stats[1].startTimer()
 		defer stats[1].stopTimer()
-		return s.commit(rec)
+		return db.s.commit(rec)
 	}, nil)
 
-	resultSize := int(int(stats[1].write))
-	s.logf("table@compaction commited F%s S%s D·%d T·%v", sint(len(rec.addedTables)-len(rec.deletedTables)), sshortenb(resultSize-sourceSize), dropCnt, stats[1].duration)
+	resultSize := int(stats[1].write)
+	db.logf("table@compaction committed F%s S%s Ke·%d D·%d T·%v", sint(len(rec.addedTables)-len(rec.deletedTables)), sshortenb(resultSize-sourceSize), b.kerrCnt, b.dropCnt, stats[1].duration)
 
 	// Save compaction stats
 	for i := range stats {
-		d.compStats[c.level+1].add(&stats[i])
+		db.compStats[c.level+1].add(&stats[i])
 	}
 }
 
-func (d *DB) tableRangeCompaction(level int, min, max []byte) {
-	s := d.s
-	s.logf("table@compaction range L%d %q:%q", level, min, max)
+func (db *DB) tableRangeCompaction(level int, umin, umax []byte) {
+	db.logf("table@compaction range L%d %q:%q", level, umin, umax)
 
 	if level >= 0 {
-		if c := s.getCompactionRange(level, min, max); c != nil {
-			d.tableCompaction(c, true)
+		if c := db.s.getCompactionRange(level, umin, umax); c != nil {
+			db.tableCompaction(c, true)
 		}
 	} else {
-		v := s.version_NB()
+		v := db.s.version()
 		m := 1
 		for i, t := range v.tables[1:] {
-			if t.isOverlaps(min, max, true, s.icmp) {
+			if t.overlaps(db.s.icmp, umin, umax, false) {
 				m = i + 1
 			}
 		}
+		v.release()
+
 		for level := 0; level < m; level++ {
-			if c := s.getCompactionRange(level, min, max); c != nil {
-				d.tableCompaction(c, true)
+			if c := db.s.getCompactionRange(level, umin, umax); c != nil {
+				db.tableCompaction(c, true)
 			}
 		}
 	}
 }
 
-func (d *DB) tableAutoCompaction() {
-	if c := d.s.pickCompaction(); c != nil {
-		d.tableCompaction(c, false)
+func (db *DB) tableAutoCompaction() {
+	if c := db.s.pickCompaction(); c != nil {
+		db.tableCompaction(c, false)
 	}
 }
 
-func (d *DB) tableNeedCompaction() bool {
-	return d.s.version_NB().needCompaction()
+func (db *DB) tableNeedCompaction() bool {
+	v := db.s.version()
+	defer v.release()
+	return v.needCompaction()
 }
 
-func (d *DB) pauseCompaction(ch chan<- struct{}) {
+func (db *DB) pauseCompaction(ch chan<- struct{}) {
 	select {
 	case ch <- struct{}{}:
-	case _, _ = <-d.closeC:
-		d.compactionExitTransact()
+	case _, _ = <-db.closeC:
+		db.compactionExitTransact()
 	}
 }
 
@@ -537,7 +665,12 @@ type cIdle struct {
 }
 
 func (r cIdle) ack(err error) {
-	r.ackC <- err
+	if r.ackC != nil {
+		defer func() {
+			recover()
+		}()
+		r.ackC <- err
+	}
 }
 
 type cRange struct {
@@ -547,56 +680,67 @@ type cRange struct {
 }
 
 func (r cRange) ack(err error) {
-	defer func() {
-		recover()
-	}()
 	if r.ackC != nil {
+		defer func() {
+			recover()
+		}()
 		r.ackC <- err
 	}
 }
 
-func (d *DB) compSendIdle(compC chan<- cCmd) error {
+// This will trigger auto compation and/or wait for all compaction to be done.
+func (db *DB) compSendIdle(compC chan<- cCmd) (err error) {
 	ch := make(chan error)
 	defer close(ch)
 	// Send cmd.
 	select {
 	case compC <- cIdle{ch}:
-	case err := <-d.compErrC:
-		return err
-	case _, _ = <-d.closeC:
+	case err = <-db.compErrC:
+		return
+	case _, _ = <-db.closeC:
 		return ErrClosed
 	}
 	// Wait cmd.
-	return <-ch
+	select {
+	case err = <-ch:
+	case err = <-db.compErrC:
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+	return err
 }
 
-func (d *DB) compSendRange(compC chan<- cCmd, level int, min, max []byte) (err error) {
+// This will trigger auto compaction but will not wait for it.
+func (db *DB) compSendTrigger(compC chan<- cCmd) {
+	select {
+	case compC <- cIdle{}:
+	default:
+	}
+}
+
+// Send range compaction request.
+func (db *DB) compSendRange(compC chan<- cCmd, level int, min, max []byte) (err error) {
 	ch := make(chan error)
 	defer close(ch)
 	// Send cmd.
 	select {
 	case compC <- cRange{level, min, max, ch}:
-	case err := <-d.compErrC:
+	case err := <-db.compErrC:
 		return err
-	case _, _ = <-d.closeC:
+	case _, _ = <-db.closeC:
 		return ErrClosed
 	}
 	// Wait cmd.
 	select {
-	case err = <-d.compErrC:
 	case err = <-ch:
+	case err = <-db.compErrC:
+	case _, _ = <-db.closeC:
+		return ErrClosed
 	}
 	return err
 }
 
-func (d *DB) compTrigger(compTriggerC chan struct{}) {
-	select {
-	case compTriggerC <- struct{}{}:
-	default:
-	}
-}
-
-func (d *DB) mCompaction() {
+func (db *DB) mCompaction() {
 	var x cCmd
 
 	defer func() {
@@ -608,24 +752,27 @@ func (d *DB) mCompaction() {
 		if x != nil {
 			x.ack(ErrClosed)
 		}
-		d.closeW.Done()
+		db.closeW.Done()
 	}()
 
 	for {
 		select {
-		case _, _ = <-d.closeC:
+		case x = <-db.mcompCmdC:
+			switch x.(type) {
+			case cIdle:
+				db.memCompaction()
+				x.ack(nil)
+				x = nil
+			default:
+				panic("leveldb: unknown command")
+			}
+		case _, _ = <-db.closeC:
 			return
-		case x = <-d.mcompCmdC:
-			d.memCompaction()
-			x.ack(nil)
-			x = nil
-		case <-d.mcompTriggerC:
-			d.memCompaction()
 		}
 	}
 }
 
-func (d *DB) tCompaction() {
+func (db *DB) tCompaction() {
 	var x cCmd
 	var ackQ []cCmd
 
@@ -642,19 +789,18 @@ func (d *DB) tCompaction() {
 		if x != nil {
 			x.ack(ErrClosed)
 		}
-		d.closeW.Done()
+		db.closeW.Done()
 	}()
 
 	for {
-		if d.tableNeedCompaction() {
+		if db.tableNeedCompaction() {
 			select {
-			case x = <-d.tcompCmdC:
-			case <-d.tcompTriggerC:
-			case _, _ = <-d.closeC:
-				return
-			case ch := <-d.tcompPauseC:
-				d.pauseCompaction(ch)
+			case x = <-db.tcompCmdC:
+			case ch := <-db.tcompPauseC:
+				db.pauseCompaction(ch)
 				continue
+			case _, _ = <-db.closeC:
+				return
 			default:
 			}
 		} else {
@@ -664,12 +810,11 @@ func (d *DB) tCompaction() {
 			}
 			ackQ = ackQ[:0]
 			select {
-			case x = <-d.tcompCmdC:
-			case <-d.tcompTriggerC:
-			case ch := <-d.tcompPauseC:
-				d.pauseCompaction(ch)
+			case x = <-db.tcompCmdC:
+			case ch := <-db.tcompPauseC:
+				db.pauseCompaction(ch)
 				continue
-			case _, _ = <-d.closeC:
+			case _, _ = <-db.closeC:
 				return
 			}
 		}
@@ -678,11 +823,13 @@ func (d *DB) tCompaction() {
 			case cIdle:
 				ackQ = append(ackQ, x)
 			case cRange:
-				d.tableRangeCompaction(cmd.level, cmd.min, cmd.max)
+				db.tableRangeCompaction(cmd.level, cmd.min, cmd.max)
 				x.ack(nil)
+			default:
+				panic("leveldb: unknown command")
 			}
 			x = nil
 		}
-		d.tableAutoCompaction()
+		db.tableAutoCompaction()
 	}
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_iter.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_iter.go
index 9973a8fef..011a94a35 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_iter.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_iter.go
@@ -8,7 +8,10 @@ package leveldb
 
 import (
 	"errors"
+	"math/rand"
 	"runtime"
+	"sync"
+	"sync/atomic"
 
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/opt"
@@ -19,50 +22,69 @@ var (
 	errInvalidIkey = errors.New("leveldb: Iterator: invalid internal key")
 )
 
-func (db *DB) newRawIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
-	s := db.s
+type memdbReleaser struct {
+	once sync.Once
+	m    *memDB
+}
+
+func (mr *memdbReleaser) Release() {
+	mr.once.Do(func() {
+		mr.m.decref()
+	})
+}
 
+func (db *DB) newRawIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
 	em, fm := db.getMems()
-	v := s.version()
+	v := db.s.version()
 
 	ti := v.getIterators(slice, ro)
 	n := len(ti) + 2
 	i := make([]iterator.Iterator, 0, n)
-	i = append(i, em.NewIterator(slice))
+	emi := em.mdb.NewIterator(slice)
+	emi.SetReleaser(&memdbReleaser{m: em})
+	i = append(i, emi)
 	if fm != nil {
-		i = append(i, fm.NewIterator(slice))
+		fmi := fm.mdb.NewIterator(slice)
+		fmi.SetReleaser(&memdbReleaser{m: fm})
+		i = append(i, fmi)
 	}
 	i = append(i, ti...)
-	strict := s.o.GetStrict(opt.StrictIterator) || ro.GetStrict(opt.StrictIterator)
-	mi := iterator.NewMergedIterator(i, s.icmp, strict)
+	strict := opt.GetStrict(db.s.o.Options, ro, opt.StrictReader)
+	mi := iterator.NewMergedIterator(i, db.s.icmp, strict)
 	mi.SetReleaser(&versionReleaser{v: v})
 	return mi
 }
 
 func (db *DB) newIterator(seq uint64, slice *util.Range, ro *opt.ReadOptions) *dbIter {
-	var slice_ *util.Range
+	var islice *util.Range
 	if slice != nil {
-		slice_ = &util.Range{}
+		islice = &util.Range{}
 		if slice.Start != nil {
-			slice_.Start = newIKey(slice.Start, kMaxSeq, tSeek)
+			islice.Start = newIkey(slice.Start, kMaxSeq, ktSeek)
 		}
 		if slice.Limit != nil {
-			slice_.Limit = newIKey(slice.Limit, kMaxSeq, tSeek)
+			islice.Limit = newIkey(slice.Limit, kMaxSeq, ktSeek)
 		}
 	}
-	rawIter := db.newRawIterator(slice_, ro)
+	rawIter := db.newRawIterator(islice, ro)
 	iter := &dbIter{
+		db:     db,
 		icmp:   db.s.icmp,
 		iter:   rawIter,
 		seq:    seq,
-		strict: db.s.o.GetStrict(opt.StrictIterator) || ro.GetStrict(opt.StrictIterator),
+		strict: opt.GetStrict(db.s.o.Options, ro, opt.StrictReader),
 		key:    make([]byte, 0),
 		value:  make([]byte, 0),
 	}
+	atomic.AddInt32(&db.aliveIters, 1)
 	runtime.SetFinalizer(iter, (*dbIter).Release)
 	return iter
 }
 
+func (db *DB) iterSamplingRate() int {
+	return rand.Intn(2 * db.s.o.GetIteratorSamplingRate())
+}
+
 type dir int
 
 const (
@@ -75,16 +97,27 @@ const (
 
 // dbIter represent an interator states over a database session.
 type dbIter struct {
+	db     *DB
 	icmp   *iComparer
 	iter   iterator.Iterator
 	seq    uint64
 	strict bool
 
-	dir      dir
-	key      []byte
-	value    []byte
-	err      error
-	releaser util.Releaser
+	smaplingGap int
+	dir         dir
+	key         []byte
+	value       []byte
+	err         error
+	releaser    util.Releaser
+}
+
+func (i *dbIter) sampleSeek() {
+	ikey := i.iter.Key()
+	i.smaplingGap -= len(ikey) + len(i.iter.Value())
+	for i.smaplingGap < 0 {
+		i.smaplingGap += i.db.iterSamplingRate()
+		i.db.sampleSeek(ikey)
+	}
 }
 
 func (i *dbIter) setErr(err error) {
@@ -144,7 +177,7 @@ func (i *dbIter) Seek(key []byte) bool {
 		return false
 	}
 
-	ikey := newIKey(key, i.seq, tSeek)
+	ikey := newIkey(key, i.seq, ktSeek)
 	if i.iter.Seek(ikey) {
 		i.dir = dirSOI
 		return i.next()
@@ -156,15 +189,15 @@ func (i *dbIter) Seek(key []byte) bool {
 
 func (i *dbIter) next() bool {
 	for {
-		ukey, seq, t, ok := parseIkey(i.iter.Key())
-		if ok {
+		if ukey, seq, kt, kerr := parseIkey(i.iter.Key()); kerr == nil {
+			i.sampleSeek()
 			if seq <= i.seq {
-				switch t {
-				case tDel:
+				switch kt {
+				case ktDel:
 					// Skip deleted key.
 					i.key = append(i.key[:0], ukey...)
 					i.dir = dirForward
-				case tVal:
+				case ktVal:
 					if i.dir == dirSOI || i.icmp.uCompare(ukey, i.key) > 0 {
 						i.key = append(i.key[:0], ukey...)
 						i.value = append(i.value[:0], i.iter.Value()...)
@@ -174,7 +207,7 @@ func (i *dbIter) next() bool {
 				}
 			}
 		} else if i.strict {
-			i.setErr(errInvalidIkey)
+			i.setErr(kerr)
 			break
 		}
 		if !i.iter.Next() {
@@ -207,20 +240,20 @@ func (i *dbIter) prev() bool {
 	del := true
 	if i.iter.Valid() {
 		for {
-			ukey, seq, t, ok := parseIkey(i.iter.Key())
-			if ok {
+			if ukey, seq, kt, kerr := parseIkey(i.iter.Key()); kerr == nil {
+				i.sampleSeek()
 				if seq <= i.seq {
 					if !del && i.icmp.uCompare(ukey, i.key) < 0 {
 						return true
 					}
-					del = (t == tDel)
+					del = (kt == ktDel)
 					if !del {
 						i.key = append(i.key[:0], ukey...)
 						i.value = append(i.value[:0], i.iter.Value()...)
 					}
 				}
 			} else if i.strict {
-				i.setErr(errInvalidIkey)
+				i.setErr(kerr)
 				return false
 			}
 			if !i.iter.Prev() {
@@ -249,13 +282,13 @@ func (i *dbIter) Prev() bool {
 		return i.Last()
 	case dirForward:
 		for i.iter.Prev() {
-			ukey, _, _, ok := parseIkey(i.iter.Key())
-			if ok {
+			if ukey, _, _, kerr := parseIkey(i.iter.Key()); kerr == nil {
+				i.sampleSeek()
 				if i.icmp.uCompare(ukey, i.key) < 0 {
 					goto cont
 				}
 			} else if i.strict {
-				i.setErr(errInvalidIkey)
+				i.setErr(kerr)
 				return false
 			}
 		}
@@ -289,6 +322,7 @@ func (i *dbIter) Release() {
 
 		if i.releaser != nil {
 			i.releaser.Release()
+			i.releaser = nil
 		}
 
 		i.dir = dirReleased
@@ -296,13 +330,19 @@ func (i *dbIter) Release() {
 		i.value = nil
 		i.iter.Release()
 		i.iter = nil
+		atomic.AddInt32(&i.db.aliveIters, -1)
+		i.db = nil
 	}
 }
 
 func (i *dbIter) SetReleaser(releaser util.Releaser) {
-	if i.dir != dirReleased {
-		i.releaser = releaser
+	if i.dir == dirReleased {
+		panic(util.ErrReleased)
+	}
+	if i.releaser != nil && releaser != nil {
+		panic(util.ErrHasReleaser)
 	}
+	i.releaser = releaser
 }
 
 func (i *dbIter) Error() error {
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_snapshot.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_snapshot.go
index 225b7cd5e..0372848ff 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_snapshot.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_snapshot.go
@@ -7,8 +7,11 @@
 package leveldb
 
 import (
+	"container/list"
+	"fmt"
 	"runtime"
 	"sync"
+	"sync/atomic"
 
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/opt"
@@ -18,51 +21,41 @@ import (
 type snapshotElement struct {
 	seq uint64
 	ref int
-	// Next and previous pointers in the doubly-linked list of elements.
-	next, prev *snapshotElement
-}
-
-// Initialize the snapshot.
-func (db *DB) initSnapshot() {
-	db.snapsRoot.next = &db.snapsRoot
-	db.snapsRoot.prev = &db.snapsRoot
+	e   *list.Element
 }
 
 // Acquires a snapshot, based on latest sequence.
 func (db *DB) acquireSnapshot() *snapshotElement {
 	db.snapsMu.Lock()
+	defer db.snapsMu.Unlock()
+
 	seq := db.getSeq()
-	elem := db.snapsRoot.prev
-	if elem == &db.snapsRoot || elem.seq != seq {
-		at := db.snapsRoot.prev
-		next := at.next
-		elem = &snapshotElement{
-			seq:  seq,
-			prev: at,
-			next: next,
+
+	if e := db.snapsList.Back(); e != nil {
+		se := e.Value.(*snapshotElement)
+		if se.seq == seq {
+			se.ref++
+			return se
+		} else if seq < se.seq {
+			panic("leveldb: sequence number is not increasing")
 		}
-		at.next = elem
-		next.prev = elem
 	}
-	elem.ref++
-	db.snapsMu.Unlock()
-	return elem
+	se := &snapshotElement{seq: seq, ref: 1}
+	se.e = db.snapsList.PushBack(se)
+	return se
 }
 
 // Releases given snapshot element.
-func (db *DB) releaseSnapshot(elem *snapshotElement) {
-	if !db.isClosed() {
-		db.snapsMu.Lock()
-		elem.ref--
-		if elem.ref == 0 {
-			elem.prev.next = elem.next
-			elem.next.prev = elem.prev
-			elem.next = nil
-			elem.prev = nil
-		} else if elem.ref < 0 {
-			panic("leveldb: Snapshot: negative element reference")
-		}
-		db.snapsMu.Unlock()
+func (db *DB) releaseSnapshot(se *snapshotElement) {
+	db.snapsMu.Lock()
+	defer db.snapsMu.Unlock()
+
+	se.ref--
+	if se.ref == 0 {
+		db.snapsList.Remove(se.e)
+		se.e = nil
+	} else if se.ref < 0 {
+		panic("leveldb: Snapshot: negative element reference")
 	}
 }
 
@@ -70,10 +63,11 @@ func (db *DB) releaseSnapshot(elem *snapshotElement) {
 func (db *DB) minSeq() uint64 {
 	db.snapsMu.Lock()
 	defer db.snapsMu.Unlock()
-	elem := db.snapsRoot.prev
-	if elem != &db.snapsRoot {
-		return elem.seq
+
+	if e := db.snapsList.Front(); e != nil {
+		return e.Value.(*snapshotElement).seq
 	}
+
 	return db.getSeq()
 }
 
@@ -81,38 +75,59 @@ func (db *DB) minSeq() uint64 {
 type Snapshot struct {
 	db       *DB
 	elem     *snapshotElement
-	mu       sync.Mutex
+	mu       sync.RWMutex
 	released bool
 }
 
 // Creates new snapshot object.
 func (db *DB) newSnapshot() *Snapshot {
-	p := &Snapshot{
+	snap := &Snapshot{
 		db:   db,
 		elem: db.acquireSnapshot(),
 	}
-	runtime.SetFinalizer(p, (*Snapshot).Release)
-	return p
+	atomic.AddInt32(&db.aliveSnaps, 1)
+	runtime.SetFinalizer(snap, (*Snapshot).Release)
+	return snap
+}
+
+func (snap *Snapshot) String() string {
+	return fmt.Sprintf("leveldb.Snapshot{%d}", snap.elem.seq)
 }
 
 // Get gets the value for the given key. It returns ErrNotFound if
-// the DB does not contain the key.
+// the DB does not contains the key.
 //
 // The caller should not modify the contents of the returned slice, but
 // it is safe to modify the contents of the argument after Get returns.
-func (p *Snapshot) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
-	db := p.db
-	err = db.ok()
+func (snap *Snapshot) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
+	err = snap.db.ok()
 	if err != nil {
 		return
 	}
-	p.mu.Lock()
-	defer p.mu.Unlock()
-	if p.released {
+	snap.mu.RLock()
+	defer snap.mu.RUnlock()
+	if snap.released {
 		err = ErrSnapshotReleased
 		return
 	}
-	return db.get(key, p.elem.seq, ro)
+	return snap.db.get(key, snap.elem.seq, ro)
+}
+
+// Has returns true if the DB does contains the given key.
+//
+// It is safe to modify the contents of the argument after Get returns.
+func (snap *Snapshot) Has(key []byte, ro *opt.ReadOptions) (ret bool, err error) {
+	err = snap.db.ok()
+	if err != nil {
+		return
+	}
+	snap.mu.RLock()
+	defer snap.mu.RUnlock()
+	if snap.released {
+		err = ErrSnapshotReleased
+		return
+	}
+	return snap.db.has(key, snap.elem.seq, ro)
 }
 
 // NewIterator returns an iterator for the snapshot of the uderlying DB.
@@ -132,17 +147,18 @@ func (p *Snapshot) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error
 // iterator would be still valid until released.
 //
 // Also read Iterator documentation of the leveldb/iterator package.
-func (p *Snapshot) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
-	db := p.db
-	if err := db.ok(); err != nil {
+func (snap *Snapshot) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
+	if err := snap.db.ok(); err != nil {
 		return iterator.NewEmptyIterator(err)
 	}
-	p.mu.Lock()
-	defer p.mu.Unlock()
-	if p.released {
+	snap.mu.Lock()
+	defer snap.mu.Unlock()
+	if snap.released {
 		return iterator.NewEmptyIterator(ErrSnapshotReleased)
 	}
-	return db.newIterator(p.elem.seq, slice, ro)
+	// Since iterator already hold version ref, it doesn't need to
+	// hold snapshot ref.
+	return snap.db.newIterator(snap.elem.seq, slice, ro)
 }
 
 // Release releases the snapshot. This will not release any returned
@@ -150,16 +166,18 @@ func (p *Snapshot) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.
 // underlying DB is closed.
 //
 // Other methods should not be called after the snapshot has been released.
-func (p *Snapshot) Release() {
-	p.mu.Lock()
-	if !p.released {
+func (snap *Snapshot) Release() {
+	snap.mu.Lock()
+	defer snap.mu.Unlock()
+
+	if !snap.released {
 		// Clear the finalizer.
-		runtime.SetFinalizer(p, nil)
+		runtime.SetFinalizer(snap, nil)
 
-		p.released = true
-		p.db.releaseSnapshot(p.elem)
-		p.db = nil
-		p.elem = nil
+		snap.released = true
+		snap.db.releaseSnapshot(snap.elem)
+		atomic.AddInt32(&snap.db.aliveSnaps, -1)
+		snap.db = nil
+		snap.elem = nil
 	}
-	p.mu.Unlock()
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_state.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_state.go
index a13706142..d4db9d6dd 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_state.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_state.go
@@ -8,106 +8,203 @@ package leveldb
 
 import (
 	"sync/atomic"
+	"time"
 
 	"github.com/syndtr/goleveldb/leveldb/journal"
 	"github.com/syndtr/goleveldb/leveldb/memdb"
 )
 
+type memDB struct {
+	db  *DB
+	mdb *memdb.DB
+	ref int32
+}
+
+func (m *memDB) incref() {
+	atomic.AddInt32(&m.ref, 1)
+}
+
+func (m *memDB) decref() {
+	if ref := atomic.AddInt32(&m.ref, -1); ref == 0 {
+		// Only put back memdb with std capacity.
+		if m.mdb.Capacity() == m.db.s.o.GetWriteBuffer() {
+			m.mdb.Reset()
+			m.db.mpoolPut(m.mdb)
+		}
+		m.db = nil
+		m.mdb = nil
+	} else if ref < 0 {
+		panic("negative memdb ref")
+	}
+}
+
 // Get latest sequence number.
-func (d *DB) getSeq() uint64 {
-	return atomic.LoadUint64(&d.seq)
+func (db *DB) getSeq() uint64 {
+	return atomic.LoadUint64(&db.seq)
 }
 
 // Atomically adds delta to seq.
-func (d *DB) addSeq(delta uint64) {
-	atomic.AddUint64(&d.seq, delta)
+func (db *DB) addSeq(delta uint64) {
+	atomic.AddUint64(&db.seq, delta)
+}
+
+func (db *DB) sampleSeek(ikey iKey) {
+	v := db.s.version()
+	if v.sampleSeek(ikey) {
+		// Trigger table compaction.
+		db.compSendTrigger(db.tcompCmdC)
+	}
+	v.release()
+}
+
+func (db *DB) mpoolPut(mem *memdb.DB) {
+	defer func() {
+		recover()
+	}()
+	select {
+	case db.memPool <- mem:
+	default:
+	}
+}
+
+func (db *DB) mpoolGet() *memdb.DB {
+	select {
+	case mem := <-db.memPool:
+		return mem
+	default:
+		return nil
+	}
+}
+
+func (db *DB) mpoolDrain() {
+	ticker := time.NewTicker(30 * time.Second)
+	for {
+		select {
+		case <-ticker.C:
+			select {
+			case <-db.memPool:
+			default:
+			}
+		case _, _ = <-db.closeC:
+			close(db.memPool)
+			return
+		}
+	}
 }
 
 // Create new memdb and froze the old one; need external synchronization.
 // newMem only called synchronously by the writer.
-func (d *DB) newMem(n int) (mem *memdb.DB, err error) {
-	s := d.s
-
-	num := s.allocFileNum()
-	file := s.getJournalFile(num)
+func (db *DB) newMem(n int) (mem *memDB, err error) {
+	num := db.s.allocFileNum()
+	file := db.s.getJournalFile(num)
 	w, err := file.Create()
 	if err != nil {
-		s.reuseFileNum(num)
+		db.s.reuseFileNum(num)
 		return
 	}
-	d.memMu.Lock()
-	if d.journal == nil {
-		d.journal = journal.NewWriter(w)
+
+	db.memMu.Lock()
+	defer db.memMu.Unlock()
+
+	if db.frozenMem != nil {
+		panic("still has frozen mem")
+	}
+
+	if db.journal == nil {
+		db.journal = journal.NewWriter(w)
 	} else {
-		d.journal.Reset(w)
-		d.journalWriter.Close()
-		d.frozenJournalFile = d.journalFile
-	}
-	d.journalWriter = w
-	d.journalFile = file
-	d.frozenMem = d.mem
-	d.mem = memdb.New(s.icmp, maxInt(d.s.o.GetWriteBuffer(), n))
-	mem = d.mem
-	// The seq only incremented by the writer.
-	d.frozenSeq = d.seq
-	d.memMu.Unlock()
+		db.journal.Reset(w)
+		db.journalWriter.Close()
+		db.frozenJournalFile = db.journalFile
+	}
+	db.journalWriter = w
+	db.journalFile = file
+	db.frozenMem = db.mem
+	mdb := db.mpoolGet()
+	if mdb == nil || mdb.Capacity() < n {
+		mdb = memdb.New(db.s.icmp, maxInt(db.s.o.GetWriteBuffer(), n))
+	}
+	mem = &memDB{
+		db:  db,
+		mdb: mdb,
+		ref: 2,
+	}
+	db.mem = mem
+	// The seq only incremented by the writer. And whoever called newMem
+	// should hold write lock, so no need additional synchronization here.
+	db.frozenSeq = db.seq
 	return
 }
 
 // Get all memdbs.
-func (d *DB) getMems() (e *memdb.DB, f *memdb.DB) {
-	d.memMu.RLock()
-	defer d.memMu.RUnlock()
-	return d.mem, d.frozenMem
+func (db *DB) getMems() (e, f *memDB) {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	if db.mem == nil {
+		panic("nil effective mem")
+	}
+	db.mem.incref()
+	if db.frozenMem != nil {
+		db.frozenMem.incref()
+	}
+	return db.mem, db.frozenMem
 }
 
 // Get frozen memdb.
-func (d *DB) getEffectiveMem() *memdb.DB {
-	d.memMu.RLock()
-	defer d.memMu.RUnlock()
-	return d.mem
+func (db *DB) getEffectiveMem() *memDB {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	if db.mem == nil {
+		panic("nil effective mem")
+	}
+	db.mem.incref()
+	return db.mem
 }
 
 // Check whether we has frozen memdb.
-func (d *DB) hasFrozenMem() bool {
-	d.memMu.RLock()
-	defer d.memMu.RUnlock()
-	return d.frozenMem != nil
+func (db *DB) hasFrozenMem() bool {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	return db.frozenMem != nil
 }
 
 // Get frozen memdb.
-func (d *DB) getFrozenMem() *memdb.DB {
-	d.memMu.RLock()
-	defer d.memMu.RUnlock()
-	return d.frozenMem
+func (db *DB) getFrozenMem() *memDB {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	if db.frozenMem != nil {
+		db.frozenMem.incref()
+	}
+	return db.frozenMem
 }
 
 // Drop frozen memdb; assume that frozen memdb isn't nil.
-func (d *DB) dropFrozenMem() {
-	d.memMu.Lock()
-	if err := d.frozenJournalFile.Remove(); err != nil {
-		d.s.logf("journal@remove removing @%d %q", d.frozenJournalFile.Num(), err)
+func (db *DB) dropFrozenMem() {
+	db.memMu.Lock()
+	if err := db.frozenJournalFile.Remove(); err != nil {
+		db.logf("journal@remove removing @%d %q", db.frozenJournalFile.Num(), err)
 	} else {
-		d.s.logf("journal@remove removed @%d", d.frozenJournalFile.Num())
+		db.logf("journal@remove removed @%d", db.frozenJournalFile.Num())
 	}
-	d.frozenJournalFile = nil
-	d.frozenMem = nil
-	d.memMu.Unlock()
+	db.frozenJournalFile = nil
+	db.frozenMem.decref()
+	db.frozenMem = nil
+	db.memMu.Unlock()
 }
 
 // Set closed flag; return true if not already closed.
-func (d *DB) setClosed() bool {
-	return atomic.CompareAndSwapUint32(&d.closed, 0, 1)
+func (db *DB) setClosed() bool {
+	return atomic.CompareAndSwapUint32(&db.closed, 0, 1)
 }
 
 // Check whether DB was closed.
-func (d *DB) isClosed() bool {
-	return atomic.LoadUint32(&d.closed) != 0
+func (db *DB) isClosed() bool {
+	return atomic.LoadUint32(&db.closed) != 0
 }
 
 // Check read ok status.
-func (d *DB) ok() error {
-	if d.isClosed() {
+func (db *DB) ok() error {
+	if db.isClosed() {
 		return ErrClosed
 	}
 	return nil
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_test.go
index 5de7d9723..38bfbf1ea 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_test.go
@@ -7,6 +7,10 @@
 package leveldb
 
 import (
+	"bytes"
+	"container/list"
+	crand "crypto/rand"
+	"encoding/binary"
 	"fmt"
 	"math/rand"
 	"os"
@@ -20,6 +24,7 @@ import (
 	"unsafe"
 
 	"github.com/syndtr/goleveldb/leveldb/comparer"
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/filter"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/opt"
@@ -148,25 +153,29 @@ func (h *dbHarness) maxNextLevelOverlappingBytes(want uint64) {
 	t := h.t
 	db := h.db
 
-	var res uint64
+	var (
+		maxOverlaps uint64
+		maxLevel    int
+	)
 	v := db.s.version()
 	for i, tt := range v.tables[1 : len(v.tables)-1] {
 		level := i + 1
 		next := v.tables[level+1]
 		for _, t := range tt {
-			var r tFiles
-			min, max := t.min.ukey(), t.max.ukey()
-			next.getOverlaps(min, max, &r, true, db.s.icmp.ucmp)
+			r := next.getOverlaps(nil, db.s.icmp, t.imin.ukey(), t.imax.ukey(), false)
 			sum := r.size()
-			if sum > res {
-				res = sum
+			if sum > maxOverlaps {
+				maxOverlaps = sum
+				maxLevel = level
 			}
 		}
 	}
 	v.release()
 
-	if res > want {
-		t.Errorf("next level overlapping bytes is more than %d, got=%d", want, res)
+	if maxOverlaps > want {
+		t.Errorf("next level most overlapping bytes is more than %d, got=%d level=%d", want, maxOverlaps, maxLevel)
+	} else {
+		t.Logf("next level most overlapping bytes is %d, level=%d want=%d", maxOverlaps, maxLevel, want)
 	}
 }
 
@@ -239,7 +248,7 @@ func (h *dbHarness) allEntriesFor(key, want string) {
 	db := h.db
 	s := db.s
 
-	ikey := newIKey([]byte(key), kMaxSeq, tVal)
+	ikey := newIkey([]byte(key), kMaxSeq, ktVal)
 	iter := db.newRawIterator(nil, nil)
 	if !iter.Seek(ikey) && iter.Error() != nil {
 		t.Error("AllEntries: error during seek, err: ", iter.Error())
@@ -248,19 +257,18 @@ func (h *dbHarness) allEntriesFor(key, want string) {
 	res := "[ "
 	first := true
 	for iter.Valid() {
-		rkey := iKey(iter.Key())
-		if _, t, ok := rkey.parseNum(); ok {
-			if s.icmp.uCompare(ikey.ukey(), rkey.ukey()) != 0 {
+		if ukey, _, kt, kerr := parseIkey(iter.Key()); kerr == nil {
+			if s.icmp.uCompare(ikey.ukey(), ukey) != 0 {
 				break
 			}
 			if !first {
 				res += ", "
 			}
 			first = false
-			switch t {
-			case tVal:
+			switch kt {
+			case ktVal:
 				res += string(iter.Value())
-			case tDel:
+			case ktDel:
 				res += "DEL"
 			}
 		} else {
@@ -325,6 +333,8 @@ func (h *dbHarness) compactMem() {
 	t := h.t
 	db := h.db
 
+	t.Log("starting memdb compaction")
+
 	db.writeLockC <- struct{}{}
 	defer func() {
 		<-db.writeLockC
@@ -340,6 +350,8 @@ func (h *dbHarness) compactMem() {
 	if h.totalTables() == 0 {
 		t.Error("zero tables after mem compaction")
 	}
+
+	t.Log("memdb compaction done")
 }
 
 func (h *dbHarness) compactRangeAtErr(level int, min, max string, wanterr bool) {
@@ -354,6 +366,8 @@ func (h *dbHarness) compactRangeAtErr(level int, min, max string, wanterr bool)
 		_max = []byte(max)
 	}
 
+	t.Logf("starting table range compaction: level=%d, min=%q, max=%q", level, min, max)
+
 	if err := db.compSendRange(db.tcompCmdC, level, _min, _max); err != nil {
 		if wanterr {
 			t.Log("CompactRangeAt: got error (expected): ", err)
@@ -363,6 +377,8 @@ func (h *dbHarness) compactRangeAtErr(level int, min, max string, wanterr bool)
 	} else if wanterr {
 		t.Error("CompactRangeAt: expect error")
 	}
+
+	t.Log("table range compaction done")
 }
 
 func (h *dbHarness) compactRangeAt(level int, min, max string) {
@@ -373,6 +389,8 @@ func (h *dbHarness) compactRange(min, max string) {
 	t := h.t
 	db := h.db
 
+	t.Logf("starting DB range compaction: min=%q, max=%q", min, max)
+
 	var r util.Range
 	if min != "" {
 		r.Start = []byte(min)
@@ -383,21 +401,25 @@ func (h *dbHarness) compactRange(min, max string) {
 	if err := db.CompactRange(r); err != nil {
 		t.Error("CompactRange: got error: ", err)
 	}
-}
 
-func (h *dbHarness) sizeAssert(start, limit string, low, hi uint64) {
-	t := h.t
-	db := h.db
+	t.Log("DB range compaction done")
+}
 
-	s, err := db.SizeOf([]util.Range{
+func (h *dbHarness) sizeOf(start, limit string) uint64 {
+	sz, err := h.db.SizeOf([]util.Range{
 		{[]byte(start), []byte(limit)},
 	})
 	if err != nil {
-		t.Error("SizeOf: got error: ", err)
+		h.t.Error("SizeOf: got error: ", err)
 	}
-	if s.Sum() < low || s.Sum() > hi {
-		t.Errorf("sizeof %q to %q not in range, want %d - %d, got %d",
-			shorten(start), shorten(limit), low, hi, s.Sum())
+	return sz.Sum()
+}
+
+func (h *dbHarness) sizeAssert(start, limit string, low, hi uint64) {
+	sz := h.sizeOf(start, limit)
+	if sz < low || sz > hi {
+		h.t.Errorf("sizeOf %q to %q not in range, want %d - %d, got %d",
+			shorten(start), shorten(limit), low, hi, sz)
 	}
 }
 
@@ -504,13 +526,13 @@ func Test_FieldsAligned(t *testing.T) {
 	p1 := new(DB)
 	testAligned(t, "DB.seq", unsafe.Offsetof(p1.seq))
 	p2 := new(session)
-	testAligned(t, "session.stFileNum", unsafe.Offsetof(p2.stFileNum))
+	testAligned(t, "session.stNextFileNum", unsafe.Offsetof(p2.stNextFileNum))
 	testAligned(t, "session.stJournalNum", unsafe.Offsetof(p2.stJournalNum))
 	testAligned(t, "session.stPrevJournalNum", unsafe.Offsetof(p2.stPrevJournalNum))
-	testAligned(t, "session.stSeq", unsafe.Offsetof(p2.stSeq))
+	testAligned(t, "session.stSeqNum", unsafe.Offsetof(p2.stSeqNum))
 }
 
-func TestDb_Locking(t *testing.T) {
+func TestDB_Locking(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.stor.Close()
 	h.openAssert(false)
@@ -518,7 +540,7 @@ func TestDb_Locking(t *testing.T) {
 	h.openAssert(true)
 }
 
-func TestDb_Empty(t *testing.T) {
+func TestDB_Empty(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.get("foo", false)
 
@@ -527,7 +549,7 @@ func TestDb_Empty(t *testing.T) {
 	})
 }
 
-func TestDb_ReadWrite(t *testing.T) {
+func TestDB_ReadWrite(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		h.getVal("foo", "v1")
@@ -542,7 +564,7 @@ func TestDb_ReadWrite(t *testing.T) {
 	})
 }
 
-func TestDb_PutDeleteGet(t *testing.T) {
+func TestDB_PutDeleteGet(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		h.getVal("foo", "v1")
@@ -556,7 +578,7 @@ func TestDb_PutDeleteGet(t *testing.T) {
 	})
 }
 
-func TestDb_EmptyBatch(t *testing.T) {
+func TestDB_EmptyBatch(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -568,7 +590,7 @@ func TestDb_EmptyBatch(t *testing.T) {
 	h.get("foo", false)
 }
 
-func TestDb_GetFromFrozen(t *testing.T) {
+func TestDB_GetFromFrozen(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{WriteBuffer: 100100})
 	defer h.close()
 
@@ -594,7 +616,7 @@ func TestDb_GetFromFrozen(t *testing.T) {
 	h.get("k2", true)
 }
 
-func TestDb_GetFromTable(t *testing.T) {
+func TestDB_GetFromTable(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		h.compactMem()
@@ -602,7 +624,7 @@ func TestDb_GetFromTable(t *testing.T) {
 	})
 }
 
-func TestDb_GetSnapshot(t *testing.T) {
+func TestDB_GetSnapshot(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		bar := strings.Repeat("b", 200)
 		h.put("foo", "v1")
@@ -636,7 +658,7 @@ func TestDb_GetSnapshot(t *testing.T) {
 	})
 }
 
-func TestDb_GetLevel0Ordering(t *testing.T) {
+func TestDB_GetLevel0Ordering(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		for i := 0; i < 4; i++ {
 			h.put("bar", fmt.Sprintf("b%d", i))
@@ -659,7 +681,7 @@ func TestDb_GetLevel0Ordering(t *testing.T) {
 	})
 }
 
-func TestDb_GetOrderedByLevels(t *testing.T) {
+func TestDB_GetOrderedByLevels(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		h.compactMem()
@@ -671,7 +693,7 @@ func TestDb_GetOrderedByLevels(t *testing.T) {
 	})
 }
 
-func TestDb_GetPicksCorrectFile(t *testing.T) {
+func TestDB_GetPicksCorrectFile(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		// Arrange to have multiple files in a non-level-0 level.
 		h.put("a", "va")
@@ -695,7 +717,7 @@ func TestDb_GetPicksCorrectFile(t *testing.T) {
 	})
 }
 
-func TestDb_GetEncountersEmptyLevel(t *testing.T) {
+func TestDB_GetEncountersEmptyLevel(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		// Arrange for the following to happen:
 		//   * sstable A in level 0
@@ -750,7 +772,7 @@ func TestDb_GetEncountersEmptyLevel(t *testing.T) {
 	})
 }
 
-func TestDb_IterMultiWithDelete(t *testing.T) {
+func TestDB_IterMultiWithDelete(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("a", "va")
 		h.put("b", "vb")
@@ -776,7 +798,7 @@ func TestDb_IterMultiWithDelete(t *testing.T) {
 	})
 }
 
-func TestDb_IteratorPinsRef(t *testing.T) {
+func TestDB_IteratorPinsRef(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -800,7 +822,7 @@ func TestDb_IteratorPinsRef(t *testing.T) {
 	iter.Release()
 }
 
-func TestDb_Recover(t *testing.T) {
+func TestDB_Recover(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		h.put("baz", "v5")
@@ -822,7 +844,7 @@ func TestDb_Recover(t *testing.T) {
 	})
 }
 
-func TestDb_RecoverWithEmptyJournal(t *testing.T) {
+func TestDB_RecoverWithEmptyJournal(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		h.put("foo", "v2")
@@ -836,7 +858,7 @@ func TestDb_RecoverWithEmptyJournal(t *testing.T) {
 	})
 }
 
-func TestDb_RecoverDuringMemtableCompaction(t *testing.T) {
+func TestDB_RecoverDuringMemtableCompaction(t *testing.T) {
 	truno(t, &opt.Options{WriteBuffer: 1000000}, func(h *dbHarness) {
 
 		h.stor.DelaySync(storage.TypeTable)
@@ -852,7 +874,7 @@ func TestDb_RecoverDuringMemtableCompaction(t *testing.T) {
 	})
 }
 
-func TestDb_MinorCompactionsHappen(t *testing.T) {
+func TestDB_MinorCompactionsHappen(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{WriteBuffer: 10000})
 	defer h.close()
 
@@ -876,7 +898,7 @@ func TestDb_MinorCompactionsHappen(t *testing.T) {
 	}
 }
 
-func TestDb_RecoverWithLargeJournal(t *testing.T) {
+func TestDB_RecoverWithLargeJournal(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -901,7 +923,7 @@ func TestDb_RecoverWithLargeJournal(t *testing.T) {
 	v.release()
 }
 
-func TestDb_CompactionsGenerateMultipleFiles(t *testing.T) {
+func TestDB_CompactionsGenerateMultipleFiles(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{
 		WriteBuffer: 10000000,
 		Compression: opt.NoCompression,
@@ -939,11 +961,11 @@ func TestDb_CompactionsGenerateMultipleFiles(t *testing.T) {
 	}
 }
 
-func TestDb_RepeatedWritesToSameKey(t *testing.T) {
+func TestDB_RepeatedWritesToSameKey(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{WriteBuffer: 100000})
 	defer h.close()
 
-	maxTables := kNumLevels + kL0_StopWritesTrigger
+	maxTables := h.o.GetNumLevel() + h.o.GetWriteL0PauseTrigger()
 
 	value := strings.Repeat("v", 2*h.o.GetWriteBuffer())
 	for i := 0; i < 5*maxTables; i++ {
@@ -955,13 +977,13 @@ func TestDb_RepeatedWritesToSameKey(t *testing.T) {
 	}
 }
 
-func TestDb_RepeatedWritesToSameKeyAfterReopen(t *testing.T) {
+func TestDB_RepeatedWritesToSameKeyAfterReopen(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{WriteBuffer: 100000})
 	defer h.close()
 
 	h.reopenDB()
 
-	maxTables := kNumLevels + kL0_StopWritesTrigger
+	maxTables := h.o.GetNumLevel() + h.o.GetWriteL0PauseTrigger()
 
 	value := strings.Repeat("v", 2*h.o.GetWriteBuffer())
 	for i := 0; i < 5*maxTables; i++ {
@@ -973,11 +995,11 @@ func TestDb_RepeatedWritesToSameKeyAfterReopen(t *testing.T) {
 	}
 }
 
-func TestDb_SparseMerge(t *testing.T) {
+func TestDB_SparseMerge(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{Compression: opt.NoCompression})
 	defer h.close()
 
-	h.putMulti(kNumLevels, "A", "Z")
+	h.putMulti(h.o.GetNumLevel(), "A", "Z")
 
 	// Suppose there is:
 	//    small amount of data with prefix A
@@ -1001,6 +1023,7 @@ func TestDb_SparseMerge(t *testing.T) {
 	h.put("C", "vc2")
 	h.compactMem()
 
+	h.waitCompaction()
 	h.maxNextLevelOverlappingBytes(20 * 1048576)
 	h.compactRangeAt(0, "", "")
 	h.waitCompaction()
@@ -1010,7 +1033,7 @@ func TestDb_SparseMerge(t *testing.T) {
 	h.maxNextLevelOverlappingBytes(20 * 1048576)
 }
 
-func TestDb_SizeOf(t *testing.T) {
+func TestDB_SizeOf(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{
 		Compression: opt.NoCompression,
 		WriteBuffer: 10000000,
@@ -1060,7 +1083,7 @@ func TestDb_SizeOf(t *testing.T) {
 	}
 }
 
-func TestDb_SizeOf_MixOfSmallAndLarge(t *testing.T) {
+func TestDB_SizeOf_MixOfSmallAndLarge(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{Compression: opt.NoCompression})
 	defer h.close()
 
@@ -1098,7 +1121,7 @@ func TestDb_SizeOf_MixOfSmallAndLarge(t *testing.T) {
 	}
 }
 
-func TestDb_Snapshot(t *testing.T) {
+func TestDB_Snapshot(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		h.put("foo", "v1")
 		s1 := h.getSnapshot()
@@ -1127,13 +1150,51 @@ func TestDb_Snapshot(t *testing.T) {
 	})
 }
 
-func TestDb_HiddenValuesAreRemoved(t *testing.T) {
+func TestDB_SnapshotList(t *testing.T) {
+	db := &DB{snapsList: list.New()}
+	e0a := db.acquireSnapshot()
+	e0b := db.acquireSnapshot()
+	db.seq = 1
+	e1 := db.acquireSnapshot()
+	db.seq = 2
+	e2 := db.acquireSnapshot()
+
+	if db.minSeq() != 0 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+	db.releaseSnapshot(e0a)
+	if db.minSeq() != 0 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+	db.releaseSnapshot(e2)
+	if db.minSeq() != 0 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+	db.releaseSnapshot(e0b)
+	if db.minSeq() != 1 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+	e2 = db.acquireSnapshot()
+	if db.minSeq() != 1 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+	db.releaseSnapshot(e1)
+	if db.minSeq() != 2 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+	db.releaseSnapshot(e2)
+	if db.minSeq() != 2 {
+		t.Fatalf("invalid sequence number, got=%d", db.minSeq())
+	}
+}
+
+func TestDB_HiddenValuesAreRemoved(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		s := h.db.s
 
 		h.put("foo", "v1")
 		h.compactMem()
-		m := kMaxMemCompactLevel
+		m := h.o.GetMaxMemCompationLevel()
 		v := s.version()
 		num := v.tLen(m)
 		v.release()
@@ -1170,14 +1231,14 @@ func TestDb_HiddenValuesAreRemoved(t *testing.T) {
 	})
 }
 
-func TestDb_DeletionMarkers2(t *testing.T) {
+func TestDB_DeletionMarkers2(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 	s := h.db.s
 
 	h.put("foo", "v1")
 	h.compactMem()
-	m := kMaxMemCompactLevel
+	m := h.o.GetMaxMemCompationLevel()
 	v := s.version()
 	num := v.tLen(m)
 	v.release()
@@ -1211,8 +1272,8 @@ func TestDb_DeletionMarkers2(t *testing.T) {
 	h.allEntriesFor("foo", "[ ]")
 }
 
-func TestDb_CompactionTableOpenError(t *testing.T) {
-	h := newDbHarnessWopt(t, &opt.Options{MaxOpenFiles: 0})
+func TestDB_CompactionTableOpenError(t *testing.T) {
+	h := newDbHarnessWopt(t, &opt.Options{OpenFilesCacheCapacity: -1})
 	defer h.close()
 
 	im := 10
@@ -1230,14 +1291,14 @@ func TestDb_CompactionTableOpenError(t *testing.T) {
 		t.Errorf("total tables is %d, want %d", n, im)
 	}
 
-	h.stor.SetOpenErr(storage.TypeTable)
+	h.stor.SetEmuErr(storage.TypeTable, tsOpOpen)
 	go h.db.CompactRange(util.Range{})
 	if err := h.db.compSendIdle(h.db.tcompCmdC); err != nil {
 		t.Log("compaction error: ", err)
 	}
 	h.closeDB0()
 	h.openDB()
-	h.stor.SetOpenErr(0)
+	h.stor.SetEmuErr(0, tsOpOpen)
 
 	for i := 0; i < im; i++ {
 		for j := 0; j < jm; j++ {
@@ -1246,9 +1307,9 @@ func TestDb_CompactionTableOpenError(t *testing.T) {
 	}
 }
 
-func TestDb_OverlapInLevel0(t *testing.T) {
+func TestDB_OverlapInLevel0(t *testing.T) {
 	trun(t, func(h *dbHarness) {
-		if kMaxMemCompactLevel != 2 {
+		if h.o.GetMaxMemCompationLevel() != 2 {
 			t.Fatal("fix test to reflect the config")
 		}
 
@@ -1289,7 +1350,7 @@ func TestDb_OverlapInLevel0(t *testing.T) {
 	})
 }
 
-func TestDb_L0_CompactionBug_Issue44_a(t *testing.T) {
+func TestDB_L0_CompactionBug_Issue44_a(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -1309,7 +1370,7 @@ func TestDb_L0_CompactionBug_Issue44_a(t *testing.T) {
 	h.getKeyVal("(a->v)")
 }
 
-func TestDb_L0_CompactionBug_Issue44_b(t *testing.T) {
+func TestDB_L0_CompactionBug_Issue44_b(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -1338,7 +1399,7 @@ func TestDb_L0_CompactionBug_Issue44_b(t *testing.T) {
 	h.getKeyVal("(->)(c->cv)")
 }
 
-func TestDb_SingleEntryMemCompaction(t *testing.T) {
+func TestDB_SingleEntryMemCompaction(t *testing.T) {
 	trun(t, func(h *dbHarness) {
 		for i := 0; i < 10; i++ {
 			h.put("big", strings.Repeat("v", opt.DefaultWriteBuffer))
@@ -1355,7 +1416,7 @@ func TestDb_SingleEntryMemCompaction(t *testing.T) {
 	})
 }
 
-func TestDb_ManifestWriteError(t *testing.T) {
+func TestDB_ManifestWriteError(t *testing.T) {
 	for i := 0; i < 2; i++ {
 		func() {
 			h := newDbHarness(t)
@@ -1368,23 +1429,23 @@ func TestDb_ManifestWriteError(t *testing.T) {
 			h.compactMem()
 			h.getVal("foo", "bar")
 			v := h.db.s.version()
-			if n := v.tLen(kMaxMemCompactLevel); n != 1 {
+			if n := v.tLen(h.o.GetMaxMemCompationLevel()); n != 1 {
 				t.Errorf("invalid total tables, want=1 got=%d", n)
 			}
 			v.release()
 
 			if i == 0 {
-				h.stor.SetWriteErr(storage.TypeManifest)
+				h.stor.SetEmuErr(storage.TypeManifest, tsOpWrite)
 			} else {
-				h.stor.SetSyncErr(storage.TypeManifest)
+				h.stor.SetEmuErr(storage.TypeManifest, tsOpSync)
 			}
 
 			// Merging compaction (will fail)
-			h.compactRangeAtErr(kMaxMemCompactLevel, "", "", true)
+			h.compactRangeAtErr(h.o.GetMaxMemCompationLevel(), "", "", true)
 
 			h.db.Close()
-			h.stor.SetWriteErr(0)
-			h.stor.SetSyncErr(0)
+			h.stor.SetEmuErr(0, tsOpWrite)
+			h.stor.SetEmuErr(0, tsOpSync)
 
 			// Should not lose data
 			h.openDB()
@@ -1405,7 +1466,7 @@ func assertErr(t *testing.T, err error, wanterr bool) {
 	}
 }
 
-func TestDb_ClosedIsClosed(t *testing.T) {
+func TestDB_ClosedIsClosed(t *testing.T) {
 	h := newDbHarness(t)
 	db := h.db
 
@@ -1500,7 +1561,7 @@ func (p numberComparer) Compare(a, b []byte) int {
 func (numberComparer) Separator(dst, a, b []byte) []byte { return nil }
 func (numberComparer) Successor(dst, b []byte) []byte    { return nil }
 
-func TestDb_CustomComparer(t *testing.T) {
+func TestDB_CustomComparer(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{
 		Comparer:    numberComparer{},
 		WriteBuffer: 1000,
@@ -1530,11 +1591,11 @@ func TestDb_CustomComparer(t *testing.T) {
 	}
 }
 
-func TestDb_ManualCompaction(t *testing.T) {
+func TestDB_ManualCompaction(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
-	if kMaxMemCompactLevel != 2 {
+	if h.o.GetMaxMemCompationLevel() != 2 {
 		t.Fatal("fix test to reflect the config")
 	}
 
@@ -1568,10 +1629,10 @@ func TestDb_ManualCompaction(t *testing.T) {
 	h.tablesPerLevel("0,0,1")
 }
 
-func TestDb_BloomFilter(t *testing.T) {
+func TestDB_BloomFilter(t *testing.T) {
 	h := newDbHarnessWopt(t, &opt.Options{
-		BlockCache: opt.NoCache,
-		Filter:     filter.NewBloomFilter(10),
+		DisableBlockCache: true,
+		Filter:            filter.NewBloomFilter(10),
 	})
 	defer h.close()
 
@@ -1579,7 +1640,7 @@ func TestDb_BloomFilter(t *testing.T) {
 		return fmt.Sprintf("key%06d", i)
 	}
 
-	n := 10000
+	const n = 10000
 
 	// Populate multiple layers
 	for i := 0; i < n; i++ {
@@ -1621,7 +1682,7 @@ func TestDb_BloomFilter(t *testing.T) {
 	h.stor.ReleaseSync(storage.TypeTable)
 }
 
-func TestDb_Concurrent(t *testing.T) {
+func TestDB_Concurrent(t *testing.T) {
 	const n, secs, maxkey = 4, 2, 1000
 
 	runtime.GOMAXPROCS(n)
@@ -1686,7 +1747,7 @@ func TestDb_Concurrent(t *testing.T) {
 	runtime.GOMAXPROCS(1)
 }
 
-func TestDb_Concurrent2(t *testing.T) {
+func TestDB_Concurrent2(t *testing.T) {
 	const n, n2 = 4, 4000
 
 	runtime.GOMAXPROCS(n*2 + 2)
@@ -1757,7 +1818,7 @@ func TestDb_Concurrent2(t *testing.T) {
 	runtime.GOMAXPROCS(1)
 }
 
-func TestDb_CreateReopenDbOnFile(t *testing.T) {
+func TestDB_CreateReopenDbOnFile(t *testing.T) {
 	dbpath := filepath.Join(os.TempDir(), fmt.Sprintf("goleveldbtestCreateReopenDbOnFile-%d", os.Getuid()))
 	if err := os.RemoveAll(dbpath); err != nil {
 		t.Fatal("cannot remove old db: ", err)
@@ -1785,7 +1846,7 @@ func TestDb_CreateReopenDbOnFile(t *testing.T) {
 	}
 }
 
-func TestDb_CreateReopenDbOnFile2(t *testing.T) {
+func TestDB_CreateReopenDbOnFile2(t *testing.T) {
 	dbpath := filepath.Join(os.TempDir(), fmt.Sprintf("goleveldbtestCreateReopenDbOnFile2-%d", os.Getuid()))
 	if err := os.RemoveAll(dbpath); err != nil {
 		t.Fatal("cannot remove old db: ", err)
@@ -1806,7 +1867,7 @@ func TestDb_CreateReopenDbOnFile2(t *testing.T) {
 	}
 }
 
-func TestDb_DeletionMarkersOnMemdb(t *testing.T) {
+func TestDB_DeletionMarkersOnMemdb(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -1817,8 +1878,8 @@ func TestDb_DeletionMarkersOnMemdb(t *testing.T) {
 	h.getKeyVal("")
 }
 
-func TestDb_LeveldbIssue178(t *testing.T) {
-	nKeys := (kMaxTableSize / 30) * 5
+func TestDB_LeveldbIssue178(t *testing.T) {
+	nKeys := (opt.DefaultCompactionTableSize / 30) * 5
 	key1 := func(i int) string {
 		return fmt.Sprintf("my_key_%d", i)
 	}
@@ -1860,7 +1921,7 @@ func TestDb_LeveldbIssue178(t *testing.T) {
 	h.assertNumKeys(nKeys)
 }
 
-func TestDb_LeveldbIssue200(t *testing.T) {
+func TestDB_LeveldbIssue200(t *testing.T) {
 	h := newDbHarness(t)
 	defer h.close()
 
@@ -1886,3 +1947,719 @@ func TestDb_LeveldbIssue200(t *testing.T) {
 	iter.Next()
 	assertBytes(t, []byte("5"), iter.Key())
 }
+
+func TestDB_GoleveldbIssue74(t *testing.T) {
+	h := newDbHarnessWopt(t, &opt.Options{
+		WriteBuffer: 1 * opt.MiB,
+	})
+	defer h.close()
+
+	const n, dur = 10000, 5 * time.Second
+
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	until := time.Now().Add(dur)
+	wg := new(sync.WaitGroup)
+	wg.Add(2)
+	var done uint32
+	go func() {
+		var i int
+		defer func() {
+			t.Logf("WRITER DONE #%d", i)
+			atomic.StoreUint32(&done, 1)
+			wg.Done()
+		}()
+
+		b := new(Batch)
+		for ; time.Now().Before(until) && atomic.LoadUint32(&done) == 0; i++ {
+			iv := fmt.Sprintf("VAL%010d", i)
+			for k := 0; k < n; k++ {
+				key := fmt.Sprintf("KEY%06d", k)
+				b.Put([]byte(key), []byte(key+iv))
+				b.Put([]byte(fmt.Sprintf("PTR%06d", k)), []byte(key))
+			}
+			h.write(b)
+
+			b.Reset()
+			snap := h.getSnapshot()
+			iter := snap.NewIterator(util.BytesPrefix([]byte("PTR")), nil)
+			var k int
+			for ; iter.Next(); k++ {
+				ptrKey := iter.Key()
+				key := iter.Value()
+
+				if _, err := snap.Get(ptrKey, nil); err != nil {
+					t.Fatalf("WRITER #%d snapshot.Get %q: %v", i, ptrKey, err)
+				}
+				if value, err := snap.Get(key, nil); err != nil {
+					t.Fatalf("WRITER #%d snapshot.Get %q: %v", i, key, err)
+				} else if string(value) != string(key)+iv {
+					t.Fatalf("WRITER #%d snapshot.Get %q got invalid value, want %q got %q", i, key, string(key)+iv, value)
+				}
+
+				b.Delete(key)
+				b.Delete(ptrKey)
+			}
+			h.write(b)
+			iter.Release()
+			snap.Release()
+			if k != n {
+				t.Fatalf("#%d %d != %d", i, k, n)
+			}
+		}
+	}()
+	go func() {
+		var i int
+		defer func() {
+			t.Logf("READER DONE #%d", i)
+			atomic.StoreUint32(&done, 1)
+			wg.Done()
+		}()
+		for ; time.Now().Before(until) && atomic.LoadUint32(&done) == 0; i++ {
+			snap := h.getSnapshot()
+			iter := snap.NewIterator(util.BytesPrefix([]byte("PTR")), nil)
+			var prevValue string
+			var k int
+			for ; iter.Next(); k++ {
+				ptrKey := iter.Key()
+				key := iter.Value()
+
+				if _, err := snap.Get(ptrKey, nil); err != nil {
+					t.Fatalf("READER #%d snapshot.Get %q: %v", i, ptrKey, err)
+				}
+
+				if value, err := snap.Get(key, nil); err != nil {
+					t.Fatalf("READER #%d snapshot.Get %q: %v", i, key, err)
+				} else if prevValue != "" && string(value) != string(key)+prevValue {
+					t.Fatalf("READER #%d snapshot.Get %q got invalid value, want %q got %q", i, key, string(key)+prevValue, value)
+				} else {
+					prevValue = string(value[len(key):])
+				}
+			}
+			iter.Release()
+			snap.Release()
+			if k > 0 && k != n {
+				t.Fatalf("#%d %d != %d", i, k, n)
+			}
+		}
+	}()
+	wg.Wait()
+}
+
+func TestDB_GetProperties(t *testing.T) {
+	h := newDbHarness(t)
+	defer h.close()
+
+	_, err := h.db.GetProperty("leveldb.num-files-at-level")
+	if err == nil {
+		t.Error("GetProperty() failed to detect missing level")
+	}
+
+	_, err = h.db.GetProperty("leveldb.num-files-at-level0")
+	if err != nil {
+		t.Error("got unexpected error", err)
+	}
+
+	_, err = h.db.GetProperty("leveldb.num-files-at-level0x")
+	if err == nil {
+		t.Error("GetProperty() failed to detect invalid level")
+	}
+}
+
+func TestDB_GoleveldbIssue72and83(t *testing.T) {
+	h := newDbHarnessWopt(t, &opt.Options{
+		WriteBuffer:            1 * opt.MiB,
+		OpenFilesCacheCapacity: 3,
+	})
+	defer h.close()
+
+	const n, wn, dur = 10000, 100, 30 * time.Second
+
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	randomData := func(prefix byte, i int) []byte {
+		data := make([]byte, 1+4+32+64+32)
+		_, err := crand.Reader.Read(data[1 : len(data)-8])
+		if err != nil {
+			panic(err)
+		}
+		data[0] = prefix
+		binary.LittleEndian.PutUint32(data[len(data)-8:], uint32(i))
+		binary.LittleEndian.PutUint32(data[len(data)-4:], util.NewCRC(data[:len(data)-4]).Value())
+		return data
+	}
+
+	keys := make([][]byte, n)
+	for i := range keys {
+		keys[i] = randomData(1, 0)
+	}
+
+	until := time.Now().Add(dur)
+	wg := new(sync.WaitGroup)
+	wg.Add(3)
+	var done uint32
+	go func() {
+		i := 0
+		defer func() {
+			t.Logf("WRITER DONE #%d", i)
+			wg.Done()
+		}()
+
+		b := new(Batch)
+		for ; i < wn && atomic.LoadUint32(&done) == 0; i++ {
+			b.Reset()
+			for _, k1 := range keys {
+				k2 := randomData(2, i)
+				b.Put(k2, randomData(42, i))
+				b.Put(k1, k2)
+			}
+			if err := h.db.Write(b, h.wo); err != nil {
+				atomic.StoreUint32(&done, 1)
+				t.Fatalf("WRITER #%d db.Write: %v", i, err)
+			}
+		}
+	}()
+	go func() {
+		var i int
+		defer func() {
+			t.Logf("READER0 DONE #%d", i)
+			atomic.StoreUint32(&done, 1)
+			wg.Done()
+		}()
+		for ; time.Now().Before(until) && atomic.LoadUint32(&done) == 0; i++ {
+			snap := h.getSnapshot()
+			seq := snap.elem.seq
+			if seq == 0 {
+				snap.Release()
+				continue
+			}
+			iter := snap.NewIterator(util.BytesPrefix([]byte{1}), nil)
+			writei := int(seq/(n*2) - 1)
+			var k int
+			for ; iter.Next(); k++ {
+				k1 := iter.Key()
+				k2 := iter.Value()
+				k1checksum0 := binary.LittleEndian.Uint32(k1[len(k1)-4:])
+				k1checksum1 := util.NewCRC(k1[:len(k1)-4]).Value()
+				if k1checksum0 != k1checksum1 {
+					t.Fatalf("READER0 #%d.%d W#%d invalid K1 checksum: %#x != %#x", i, k, k1checksum0, k1checksum0)
+				}
+				k2checksum0 := binary.LittleEndian.Uint32(k2[len(k2)-4:])
+				k2checksum1 := util.NewCRC(k2[:len(k2)-4]).Value()
+				if k2checksum0 != k2checksum1 {
+					t.Fatalf("READER0 #%d.%d W#%d invalid K2 checksum: %#x != %#x", i, k, k2checksum0, k2checksum1)
+				}
+				kwritei := int(binary.LittleEndian.Uint32(k2[len(k2)-8:]))
+				if writei != kwritei {
+					t.Fatalf("READER0 #%d.%d W#%d invalid write iteration num: %d", i, k, writei, kwritei)
+				}
+				if _, err := snap.Get(k2, nil); err != nil {
+					t.Fatalf("READER0 #%d.%d W#%d snap.Get: %v\nk1: %x\n -> k2: %x", i, k, writei, err, k1, k2)
+				}
+			}
+			if err := iter.Error(); err != nil {
+				t.Fatalf("READER0 #%d.%d W#%d snap.Iterator: %v", i, k, writei, err)
+			}
+			iter.Release()
+			snap.Release()
+			if k > 0 && k != n {
+				t.Fatalf("READER0 #%d W#%d short read, got=%d want=%d", i, writei, k, n)
+			}
+		}
+	}()
+	go func() {
+		var i int
+		defer func() {
+			t.Logf("READER1 DONE #%d", i)
+			atomic.StoreUint32(&done, 1)
+			wg.Done()
+		}()
+		for ; time.Now().Before(until) && atomic.LoadUint32(&done) == 0; i++ {
+			iter := h.db.NewIterator(nil, nil)
+			seq := iter.(*dbIter).seq
+			if seq == 0 {
+				iter.Release()
+				continue
+			}
+			writei := int(seq/(n*2) - 1)
+			var k int
+			for ok := iter.Last(); ok; ok = iter.Prev() {
+				k++
+			}
+			if err := iter.Error(); err != nil {
+				t.Fatalf("READER1 #%d.%d W#%d db.Iterator: %v", i, k, writei, err)
+			}
+			iter.Release()
+			if m := (writei+1)*n + n; k != m {
+				t.Fatalf("READER1 #%d W#%d short read, got=%d want=%d", i, writei, k, m)
+			}
+		}
+	}()
+
+	wg.Wait()
+}
+
+func TestDB_TransientError(t *testing.T) {
+	h := newDbHarnessWopt(t, &opt.Options{
+		WriteBuffer:              128 * opt.KiB,
+		OpenFilesCacheCapacity:   3,
+		DisableCompactionBackoff: true,
+	})
+	defer h.close()
+
+	const (
+		nSnap = 20
+		nKey  = 10000
+	)
+
+	var (
+		snaps [nSnap]*Snapshot
+		b     = &Batch{}
+	)
+	for i := range snaps {
+		vtail := fmt.Sprintf("VAL%030d", i)
+		b.Reset()
+		for k := 0; k < nKey; k++ {
+			key := fmt.Sprintf("KEY%8d", k)
+			b.Put([]byte(key), []byte(key+vtail))
+		}
+		h.stor.SetEmuRandErr(storage.TypeTable, tsOpOpen, tsOpRead, tsOpReadAt)
+		if err := h.db.Write(b, nil); err != nil {
+			t.Logf("WRITE #%d error: %v", i, err)
+			h.stor.SetEmuRandErr(0, tsOpOpen, tsOpRead, tsOpReadAt, tsOpWrite)
+			for {
+				if err := h.db.Write(b, nil); err == nil {
+					break
+				} else if errors.IsCorrupted(err) {
+					t.Fatalf("WRITE #%d corrupted: %v", i, err)
+				}
+			}
+		}
+
+		snaps[i] = h.db.newSnapshot()
+		b.Reset()
+		for k := 0; k < nKey; k++ {
+			key := fmt.Sprintf("KEY%8d", k)
+			b.Delete([]byte(key))
+		}
+		h.stor.SetEmuRandErr(storage.TypeTable, tsOpOpen, tsOpRead, tsOpReadAt)
+		if err := h.db.Write(b, nil); err != nil {
+			t.Logf("WRITE #%d  error: %v", i, err)
+			h.stor.SetEmuRandErr(0, tsOpOpen, tsOpRead, tsOpReadAt)
+			for {
+				if err := h.db.Write(b, nil); err == nil {
+					break
+				} else if errors.IsCorrupted(err) {
+					t.Fatalf("WRITE #%d corrupted: %v", i, err)
+				}
+			}
+		}
+	}
+	h.stor.SetEmuRandErr(0, tsOpOpen, tsOpRead, tsOpReadAt)
+
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	rnd := rand.New(rand.NewSource(0xecafdaed))
+	wg := &sync.WaitGroup{}
+	for i, snap := range snaps {
+		wg.Add(2)
+
+		go func(i int, snap *Snapshot, sk []int) {
+			defer wg.Done()
+
+			vtail := fmt.Sprintf("VAL%030d", i)
+			for _, k := range sk {
+				key := fmt.Sprintf("KEY%8d", k)
+				xvalue, err := snap.Get([]byte(key), nil)
+				if err != nil {
+					t.Fatalf("READER_GET #%d SEQ=%d K%d error: %v", i, snap.elem.seq, k, err)
+				}
+				value := key + vtail
+				if !bytes.Equal([]byte(value), xvalue) {
+					t.Fatalf("READER_GET #%d SEQ=%d K%d invalid value: want %q, got %q", i, snap.elem.seq, k, value, xvalue)
+				}
+			}
+		}(i, snap, rnd.Perm(nKey))
+
+		go func(i int, snap *Snapshot) {
+			defer wg.Done()
+
+			vtail := fmt.Sprintf("VAL%030d", i)
+			iter := snap.NewIterator(nil, nil)
+			defer iter.Release()
+			for k := 0; k < nKey; k++ {
+				if !iter.Next() {
+					if err := iter.Error(); err != nil {
+						t.Fatalf("READER_ITER #%d K%d error: %v", i, k, err)
+					} else {
+						t.Fatalf("READER_ITER #%d K%d eoi", i, k)
+					}
+				}
+				key := fmt.Sprintf("KEY%8d", k)
+				xkey := iter.Key()
+				if !bytes.Equal([]byte(key), xkey) {
+					t.Fatalf("READER_ITER #%d K%d invalid key: want %q, got %q", i, k, key, xkey)
+				}
+				value := key + vtail
+				xvalue := iter.Value()
+				if !bytes.Equal([]byte(value), xvalue) {
+					t.Fatalf("READER_ITER #%d K%d invalid value: want %q, got %q", i, k, value, xvalue)
+				}
+			}
+		}(i, snap)
+	}
+
+	wg.Wait()
+}
+
+func TestDB_UkeyShouldntHopAcrossTable(t *testing.T) {
+	h := newDbHarnessWopt(t, &opt.Options{
+		WriteBuffer:                 112 * opt.KiB,
+		CompactionTableSize:         90 * opt.KiB,
+		CompactionExpandLimitFactor: 1,
+	})
+	defer h.close()
+
+	const (
+		nSnap = 190
+		nKey  = 140
+	)
+
+	var (
+		snaps [nSnap]*Snapshot
+		b     = &Batch{}
+	)
+	for i := range snaps {
+		vtail := fmt.Sprintf("VAL%030d", i)
+		b.Reset()
+		for k := 0; k < nKey; k++ {
+			key := fmt.Sprintf("KEY%08d", k)
+			b.Put([]byte(key), []byte(key+vtail))
+		}
+		if err := h.db.Write(b, nil); err != nil {
+			t.Fatalf("WRITE #%d error: %v", i, err)
+		}
+
+		snaps[i] = h.db.newSnapshot()
+		b.Reset()
+		for k := 0; k < nKey; k++ {
+			key := fmt.Sprintf("KEY%08d", k)
+			b.Delete([]byte(key))
+		}
+		if err := h.db.Write(b, nil); err != nil {
+			t.Fatalf("WRITE #%d  error: %v", i, err)
+		}
+	}
+
+	h.compactMem()
+
+	h.waitCompaction()
+	for level, tables := range h.db.s.stVersion.tables {
+		for _, table := range tables {
+			t.Logf("L%d@%d %q:%q", level, table.file.Num(), table.imin, table.imax)
+		}
+	}
+
+	h.compactRangeAt(0, "", "")
+	h.waitCompaction()
+	for level, tables := range h.db.s.stVersion.tables {
+		for _, table := range tables {
+			t.Logf("L%d@%d %q:%q", level, table.file.Num(), table.imin, table.imax)
+		}
+	}
+	h.compactRangeAt(1, "", "")
+	h.waitCompaction()
+	for level, tables := range h.db.s.stVersion.tables {
+		for _, table := range tables {
+			t.Logf("L%d@%d %q:%q", level, table.file.Num(), table.imin, table.imax)
+		}
+	}
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	wg := &sync.WaitGroup{}
+	for i, snap := range snaps {
+		wg.Add(1)
+
+		go func(i int, snap *Snapshot) {
+			defer wg.Done()
+
+			vtail := fmt.Sprintf("VAL%030d", i)
+			for k := 0; k < nKey; k++ {
+				key := fmt.Sprintf("KEY%08d", k)
+				xvalue, err := snap.Get([]byte(key), nil)
+				if err != nil {
+					t.Fatalf("READER_GET #%d SEQ=%d K%d error: %v", i, snap.elem.seq, k, err)
+				}
+				value := key + vtail
+				if !bytes.Equal([]byte(value), xvalue) {
+					t.Fatalf("READER_GET #%d SEQ=%d K%d invalid value: want %q, got %q", i, snap.elem.seq, k, value, xvalue)
+				}
+			}
+		}(i, snap)
+	}
+
+	wg.Wait()
+}
+
+func TestDB_TableCompactionBuilder(t *testing.T) {
+	stor := newTestStorage(t)
+	defer stor.Close()
+
+	const nSeq = 99
+
+	o := &opt.Options{
+		WriteBuffer:                 112 * opt.KiB,
+		CompactionTableSize:         43 * opt.KiB,
+		CompactionExpandLimitFactor: 1,
+		CompactionGPOverlapsFactor:  1,
+		DisableBlockCache:           true,
+	}
+	s, err := newSession(stor, o)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := s.create(); err != nil {
+		t.Fatal(err)
+	}
+	defer s.close()
+	var (
+		seq        uint64
+		targetSize = 5 * o.CompactionTableSize
+		value      = bytes.Repeat([]byte{'0'}, 100)
+	)
+	for i := 0; i < 2; i++ {
+		tw, err := s.tops.create()
+		if err != nil {
+			t.Fatal(err)
+		}
+		for k := 0; tw.tw.BytesLen() < targetSize; k++ {
+			key := []byte(fmt.Sprintf("%09d", k))
+			seq += nSeq - 1
+			for x := uint64(0); x < nSeq; x++ {
+				if err := tw.append(newIkey(key, seq-x, ktVal), value); err != nil {
+					t.Fatal(err)
+				}
+			}
+		}
+		tf, err := tw.finish()
+		if err != nil {
+			t.Fatal(err)
+		}
+		rec := &sessionRecord{numLevel: s.o.GetNumLevel()}
+		rec.addTableFile(i, tf)
+		if err := s.commit(rec); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	// Build grandparent.
+	v := s.version()
+	c := newCompaction(s, v, 1, append(tFiles{}, v.tables[1]...))
+	rec := &sessionRecord{numLevel: s.o.GetNumLevel()}
+	b := &tableCompactionBuilder{
+		s:         s,
+		c:         c,
+		rec:       rec,
+		stat1:     new(cStatsStaging),
+		minSeq:    0,
+		strict:    true,
+		tableSize: o.CompactionTableSize/3 + 961,
+	}
+	if err := b.run(new(compactionTransactCounter)); err != nil {
+		t.Fatal(err)
+	}
+	for _, t := range c.tables[0] {
+		rec.delTable(c.level, t.file.Num())
+	}
+	if err := s.commit(rec); err != nil {
+		t.Fatal(err)
+	}
+	c.release()
+
+	// Build level-1.
+	v = s.version()
+	c = newCompaction(s, v, 0, append(tFiles{}, v.tables[0]...))
+	rec = &sessionRecord{numLevel: s.o.GetNumLevel()}
+	b = &tableCompactionBuilder{
+		s:         s,
+		c:         c,
+		rec:       rec,
+		stat1:     new(cStatsStaging),
+		minSeq:    0,
+		strict:    true,
+		tableSize: o.CompactionTableSize,
+	}
+	if err := b.run(new(compactionTransactCounter)); err != nil {
+		t.Fatal(err)
+	}
+	for _, t := range c.tables[0] {
+		rec.delTable(c.level, t.file.Num())
+	}
+	// Move grandparent to level-3
+	for _, t := range v.tables[2] {
+		rec.delTable(2, t.file.Num())
+		rec.addTableFile(3, t)
+	}
+	if err := s.commit(rec); err != nil {
+		t.Fatal(err)
+	}
+	c.release()
+
+	v = s.version()
+	for level, want := range []bool{false, true, false, true, false} {
+		got := len(v.tables[level]) > 0
+		if want != got {
+			t.Fatalf("invalid level-%d tables len: want %v, got %v", level, want, got)
+		}
+	}
+	for i, f := range v.tables[1][:len(v.tables[1])-1] {
+		nf := v.tables[1][i+1]
+		if bytes.Equal(f.imax.ukey(), nf.imin.ukey()) {
+			t.Fatalf("KEY %q hop across table %d .. %d", f.imax.ukey(), f.file.Num(), nf.file.Num())
+		}
+	}
+	v.release()
+
+	// Compaction with transient error.
+	v = s.version()
+	c = newCompaction(s, v, 1, append(tFiles{}, v.tables[1]...))
+	rec = &sessionRecord{numLevel: s.o.GetNumLevel()}
+	b = &tableCompactionBuilder{
+		s:         s,
+		c:         c,
+		rec:       rec,
+		stat1:     new(cStatsStaging),
+		minSeq:    0,
+		strict:    true,
+		tableSize: o.CompactionTableSize,
+	}
+	stor.SetEmuErrOnce(storage.TypeTable, tsOpSync)
+	stor.SetEmuRandErr(storage.TypeTable, tsOpRead, tsOpReadAt, tsOpWrite)
+	stor.SetEmuRandErrProb(0xf0)
+	for {
+		if err := b.run(new(compactionTransactCounter)); err != nil {
+			t.Logf("(expected) b.run: %v", err)
+		} else {
+			break
+		}
+	}
+	if err := s.commit(rec); err != nil {
+		t.Fatal(err)
+	}
+	c.release()
+
+	stor.SetEmuErrOnce(0, tsOpSync)
+	stor.SetEmuRandErr(0, tsOpRead, tsOpReadAt, tsOpWrite)
+
+	v = s.version()
+	if len(v.tables[1]) != len(v.tables[2]) {
+		t.Fatalf("invalid tables length, want %d, got %d", len(v.tables[1]), len(v.tables[2]))
+	}
+	for i, f0 := range v.tables[1] {
+		f1 := v.tables[2][i]
+		iter0 := s.tops.newIterator(f0, nil, nil)
+		iter1 := s.tops.newIterator(f1, nil, nil)
+		for j := 0; true; j++ {
+			next0 := iter0.Next()
+			next1 := iter1.Next()
+			if next0 != next1 {
+				t.Fatalf("#%d.%d invalid eoi: want %v, got %v", i, j, next0, next1)
+			}
+			key0 := iter0.Key()
+			key1 := iter1.Key()
+			if !bytes.Equal(key0, key1) {
+				t.Fatalf("#%d.%d invalid key: want %q, got %q", i, j, key0, key1)
+			}
+			if next0 == false {
+				break
+			}
+		}
+		iter0.Release()
+		iter1.Release()
+	}
+	v.release()
+}
+
+func testDB_IterTriggeredCompaction(t *testing.T, limitDiv int) {
+	const (
+		vSize = 200 * opt.KiB
+		tSize = 100 * opt.MiB
+		mIter = 100
+		n     = tSize / vSize
+	)
+
+	h := newDbHarnessWopt(t, &opt.Options{
+		Compression:       opt.NoCompression,
+		DisableBlockCache: true,
+	})
+	defer h.close()
+
+	key := func(x int) string {
+		return fmt.Sprintf("v%06d", x)
+	}
+
+	// Fill.
+	value := strings.Repeat("x", vSize)
+	for i := 0; i < n; i++ {
+		h.put(key(i), value)
+	}
+	h.compactMem()
+
+	// Delete all.
+	for i := 0; i < n; i++ {
+		h.delete(key(i))
+	}
+	h.compactMem()
+
+	var (
+		limit = n / limitDiv
+
+		startKey = key(0)
+		limitKey = key(limit)
+		maxKey   = key(n)
+		slice    = &util.Range{Limit: []byte(limitKey)}
+
+		initialSize0 = h.sizeOf(startKey, limitKey)
+		initialSize1 = h.sizeOf(limitKey, maxKey)
+	)
+
+	t.Logf("inital size %s [rest %s]", shortenb(int(initialSize0)), shortenb(int(initialSize1)))
+
+	for r := 0; true; r++ {
+		if r >= mIter {
+			t.Fatal("taking too long to compact")
+		}
+
+		// Iterates.
+		iter := h.db.NewIterator(slice, h.ro)
+		for iter.Next() {
+		}
+		if err := iter.Error(); err != nil {
+			t.Fatalf("Iter err: %v", err)
+		}
+		iter.Release()
+
+		// Wait compaction.
+		h.waitCompaction()
+
+		// Check size.
+		size0 := h.sizeOf(startKey, limitKey)
+		size1 := h.sizeOf(limitKey, maxKey)
+		t.Logf("#%03d size %s [rest %s]", r, shortenb(int(size0)), shortenb(int(size1)))
+		if size0 < initialSize0/10 {
+			break
+		}
+	}
+
+	if initialSize1 > 0 {
+		h.sizeAssert(limitKey, maxKey, initialSize1/4-opt.MiB, initialSize1+opt.MiB)
+	}
+}
+
+func TestDB_IterTriggeredCompaction(t *testing.T) {
+	testDB_IterTriggeredCompaction(t, 1)
+}
+
+func TestDB_IterTriggeredCompactionHalf(t *testing.T) {
+	testDB_IterTriggeredCompaction(t, 2)
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_util.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_util.go
index 2db007e27..a8a2bdf72 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_util.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_util.go
@@ -7,8 +7,7 @@
 package leveldb
 
 import (
-	"errors"
-
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/opt"
 	"github.com/syndtr/goleveldb/leveldb/storage"
@@ -32,40 +31,44 @@ func (p Sizes) Sum() (n uint64) {
 	return n
 }
 
+// Logging.
+func (db *DB) log(v ...interface{})                 { db.s.log(v...) }
+func (db *DB) logf(format string, v ...interface{}) { db.s.logf(format, v...) }
+
 // Check and clean files.
-func (d *DB) checkAndCleanFiles() error {
-	s := d.s
+func (db *DB) checkAndCleanFiles() error {
+	v := db.s.version()
+	defer v.release()
 
-	v := s.version_NB()
-	tables := make(map[uint64]bool)
-	for _, tt := range v.tables {
-		for _, t := range tt {
-			tables[t.file.Num()] = false
+	tablesMap := make(map[uint64]bool)
+	for _, tables := range v.tables {
+		for _, t := range tables {
+			tablesMap[t.file.Num()] = false
 		}
 	}
 
-	ff, err := s.getFiles(storage.TypeAll)
+	files, err := db.s.getFiles(storage.TypeAll)
 	if err != nil {
 		return err
 	}
 
 	var nTables int
 	var rem []storage.File
-	for _, f := range ff {
+	for _, f := range files {
 		keep := true
 		switch f.Type() {
 		case storage.TypeManifest:
-			keep = f.Num() >= s.manifestFile.Num()
+			keep = f.Num() >= db.s.manifestFile.Num()
 		case storage.TypeJournal:
-			if d.frozenJournalFile != nil {
-				keep = f.Num() >= d.frozenJournalFile.Num()
+			if db.frozenJournalFile != nil {
+				keep = f.Num() >= db.frozenJournalFile.Num()
 			} else {
-				keep = f.Num() >= d.journalFile.Num()
+				keep = f.Num() >= db.journalFile.Num()
 			}
 		case storage.TypeTable:
-			_, keep = tables[f.Num()]
+			_, keep = tablesMap[f.Num()]
 			if keep {
-				tables[f.Num()] = true
+				tablesMap[f.Num()] = true
 				nTables++
 			}
 		}
@@ -75,18 +78,20 @@ func (d *DB) checkAndCleanFiles() error {
 		}
 	}
 
-	if nTables != len(tables) {
-		for num, present := range tables {
+	if nTables != len(tablesMap) {
+		var missing []*storage.FileInfo
+		for num, present := range tablesMap {
 			if !present {
-				s.logf("db@janitor table missing @%d", num)
+				missing = append(missing, &storage.FileInfo{Type: storage.TypeTable, Num: num})
+				db.logf("db@janitor table missing @%d", num)
 			}
 		}
-		return ErrCorrupted{Type: MissingFiles, Err: errors.New("leveldb: table files missing")}
+		return errors.NewErrCorrupted(nil, &errors.ErrMissingFiles{Files: missing})
 	}
 
-	s.logf("db@janitor F·%d G·%d", len(ff), len(rem))
+	db.logf("db@janitor F·%d G·%d", len(files), len(rem))
 	for _, f := range rem {
-		s.logf("db@janitor removing %s-%d", f.Type(), f.Num())
+		db.logf("db@janitor removing %s-%d", f.Type(), f.Num())
 		if err := f.Remove(); err != nil {
 			return err
 		}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_write.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_write.go
index 4660e840c..e1cf30c53 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_write.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/db_write.go
@@ -14,84 +14,93 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-func (d *DB) writeJournal(b *Batch) error {
-	w, err := d.journal.Next()
+func (db *DB) writeJournal(b *Batch) error {
+	w, err := db.journal.Next()
 	if err != nil {
 		return err
 	}
 	if _, err := w.Write(b.encode()); err != nil {
 		return err
 	}
-	if err := d.journal.Flush(); err != nil {
+	if err := db.journal.Flush(); err != nil {
 		return err
 	}
 	if b.sync {
-		return d.journalWriter.Sync()
+		return db.journalWriter.Sync()
 	}
 	return nil
 }
 
-func (d *DB) jWriter() {
-	defer d.closeW.Done()
+func (db *DB) jWriter() {
+	defer db.closeW.Done()
 	for {
 		select {
-		case b := <-d.journalC:
+		case b := <-db.journalC:
 			if b != nil {
-				d.journalAckC <- d.writeJournal(b)
+				db.journalAckC <- db.writeJournal(b)
 			}
-		case _, _ = <-d.closeC:
+		case _, _ = <-db.closeC:
 			return
 		}
 	}
 }
 
-func (d *DB) rotateMem(n int) (mem *memdb.DB, err error) {
+func (db *DB) rotateMem(n int) (mem *memDB, err error) {
 	// Wait for pending memdb compaction.
-	err = d.compSendIdle(d.mcompCmdC)
+	err = db.compSendIdle(db.mcompCmdC)
 	if err != nil {
 		return
 	}
 
 	// Create new memdb and journal.
-	mem, err = d.newMem(n)
+	mem, err = db.newMem(n)
 	if err != nil {
 		return
 	}
 
 	// Schedule memdb compaction.
-	d.compTrigger(d.mcompTriggerC)
+	db.compSendTrigger(db.mcompCmdC)
 	return
 }
 
-func (d *DB) flush(n int) (mem *memdb.DB, nn int, err error) {
-	s := d.s
-
+func (db *DB) flush(n int) (mem *memDB, nn int, err error) {
 	delayed := false
-	flush := func() bool {
-		v := s.version()
+	flush := func() (retry bool) {
+		v := db.s.version()
 		defer v.release()
-		mem = d.getEffectiveMem()
-		nn = mem.Free()
+		mem = db.getEffectiveMem()
+		defer func() {
+			if retry {
+				mem.decref()
+				mem = nil
+			}
+		}()
+		nn = mem.mdb.Free()
 		switch {
-		case v.tLen(0) >= kL0_SlowdownWritesTrigger && !delayed:
+		case v.tLen(0) >= db.s.o.GetWriteL0SlowdownTrigger() && !delayed:
 			delayed = true
 			time.Sleep(time.Millisecond)
 		case nn >= n:
 			return false
-		case v.tLen(0) >= kL0_StopWritesTrigger:
+		case v.tLen(0) >= db.s.o.GetWriteL0PauseTrigger():
 			delayed = true
-			err = d.compSendIdle(d.tcompCmdC)
+			err = db.compSendIdle(db.tcompCmdC)
 			if err != nil {
 				return false
 			}
 		default:
 			// Allow memdb to grow if it has no entry.
-			if mem.Len() == 0 {
+			if mem.mdb.Len() == 0 {
 				nn = n
-				return false
+			} else {
+				mem.decref()
+				mem, err = db.rotateMem(n)
+				if err == nil {
+					nn = mem.mdb.Free()
+				} else {
+					nn = 0
+				}
 			}
-			mem, err = d.rotateMem(n)
-			nn = mem.Free()
 			return false
 		}
 		return true
@@ -100,7 +109,12 @@ func (d *DB) flush(n int) (mem *memdb.DB, nn int, err error) {
 	for flush() {
 	}
 	if delayed {
-		s.logf("db@write delayed T·%v", time.Since(start))
+		db.writeDelay += time.Since(start)
+		db.writeDelayN++
+	} else if db.writeDelayN > 0 {
+		db.logf("db@write was delayed N·%d T·%v", db.writeDelayN, db.writeDelay)
+		db.writeDelay = 0
+		db.writeDelayN = 0
 	}
 	return
 }
@@ -109,39 +123,45 @@ func (d *DB) flush(n int) (mem *memdb.DB, nn int, err error) {
 // sequentially.
 //
 // It is safe to modify the contents of the arguments after Write returns.
-func (d *DB) Write(b *Batch, wo *opt.WriteOptions) (err error) {
-	err = d.ok()
-	if err != nil || b == nil || b.len() == 0 {
+func (db *DB) Write(b *Batch, wo *opt.WriteOptions) (err error) {
+	err = db.ok()
+	if err != nil || b == nil || b.Len() == 0 {
 		return
 	}
 
 	b.init(wo.GetSync())
 
 	// The write happen synchronously.
-retry:
 	select {
-	case d.writeC <- b:
-		if <-d.writeMergedC {
-			return <-d.writeAckC
+	case db.writeC <- b:
+		if <-db.writeMergedC {
+			return <-db.writeAckC
 		}
-		goto retry
-	case d.writeLockC <- struct{}{}:
-	case _, _ = <-d.closeC:
+	case db.writeLockC <- struct{}{}:
+	case err = <-db.compPerErrC:
+		return
+	case _, _ = <-db.closeC:
 		return ErrClosed
 	}
 
 	merged := 0
+	danglingMerge := false
 	defer func() {
-		<-d.writeLockC
+		if danglingMerge {
+			db.writeMergedC <- false
+		} else {
+			<-db.writeLockC
+		}
 		for i := 0; i < merged; i++ {
-			d.writeAckC <- err
+			db.writeAckC <- err
 		}
 	}()
 
-	mem, memFree, err := d.flush(b.size())
+	mem, memFree, err := db.flush(b.size())
 	if err != nil {
 		return
 	}
+	defer mem.decref()
 
 	// Calculate maximum size of the batch.
 	m := 1 << 20
@@ -154,13 +174,13 @@ retry:
 drain:
 	for b.size() < m && !b.sync {
 		select {
-		case nb := <-d.writeC:
+		case nb := <-db.writeC:
 			if b.size()+nb.size() <= m {
 				b.append(nb)
-				d.writeMergedC <- true
+				db.writeMergedC <- true
 				merged++
 			} else {
-				d.writeMergedC <- false
+				danglingMerge = true
 				break drain
 			}
 		default:
@@ -169,44 +189,52 @@ drain:
 	}
 
 	// Set batch first seq number relative from last seq.
-	b.seq = d.seq + 1
+	b.seq = db.seq + 1
 
 	// Write journal concurrently if it is large enough.
 	if b.size() >= (128 << 10) {
 		// Push the write batch to the journal writer
 		select {
-		case _, _ = <-d.closeC:
+		case db.journalC <- b:
+			// Write into memdb
+			if berr := b.memReplay(mem.mdb); berr != nil {
+				panic(berr)
+			}
+		case err = <-db.compPerErrC:
+			return
+		case _, _ = <-db.closeC:
 			err = ErrClosed
 			return
-		case d.journalC <- b:
-			// Write into memdb
-			b.memReplay(mem)
 		}
 		// Wait for journal writer
 		select {
-		case _, _ = <-d.closeC:
-			err = ErrClosed
-			return
-		case err = <-d.journalAckC:
+		case err = <-db.journalAckC:
 			if err != nil {
 				// Revert memdb if error detected
-				b.revertMemReplay(mem)
+				if berr := b.revertMemReplay(mem.mdb); berr != nil {
+					panic(berr)
+				}
 				return
 			}
+		case _, _ = <-db.closeC:
+			err = ErrClosed
+			return
 		}
 	} else {
-		err = d.writeJournal(b)
+		err = db.writeJournal(b)
 		if err != nil {
 			return
 		}
-		b.memReplay(mem)
+		if berr := b.memReplay(mem.mdb); berr != nil {
+			panic(berr)
+		}
 	}
 
 	// Set last seq number.
-	d.addSeq(uint64(b.len()))
+	db.addSeq(uint64(b.Len()))
 
 	if b.size() >= memFree {
-		d.rotateMem(0)
+		db.rotateMem(0)
 	}
 	return
 }
@@ -215,20 +243,20 @@ drain:
 // for that key; a DB is not a multi-map.
 //
 // It is safe to modify the contents of the arguments after Put returns.
-func (d *DB) Put(key, value []byte, wo *opt.WriteOptions) error {
+func (db *DB) Put(key, value []byte, wo *opt.WriteOptions) error {
 	b := new(Batch)
 	b.Put(key, value)
-	return d.Write(b, wo)
+	return db.Write(b, wo)
 }
 
 // Delete deletes the value for the given key. It returns ErrNotFound if
 // the DB does not contain the key.
 //
 // It is safe to modify the contents of the arguments after Delete returns.
-func (d *DB) Delete(key []byte, wo *opt.WriteOptions) error {
+func (db *DB) Delete(key []byte, wo *opt.WriteOptions) error {
 	b := new(Batch)
 	b.Delete(key)
-	return d.Write(b, wo)
+	return db.Write(b, wo)
 }
 
 func isMemOverlaps(icmp *iComparer, mem *memdb.DB, min, max []byte) bool {
@@ -247,33 +275,37 @@ func isMemOverlaps(icmp *iComparer, mem *memdb.DB, min, max []byte) bool {
 // A nil Range.Start is treated as a key before all keys in the DB.
 // And a nil Range.Limit is treated as a key after all keys in the DB.
 // Therefore if both is nil then it will compact entire DB.
-func (d *DB) CompactRange(r util.Range) error {
-	if err := d.ok(); err != nil {
+func (db *DB) CompactRange(r util.Range) error {
+	if err := db.ok(); err != nil {
 		return err
 	}
 
+	// Lock writer.
 	select {
-	case d.writeLockC <- struct{}{}:
-	case _, _ = <-d.closeC:
+	case db.writeLockC <- struct{}{}:
+	case err := <-db.compPerErrC:
+		return err
+	case _, _ = <-db.closeC:
 		return ErrClosed
 	}
 
 	// Check for overlaps in memdb.
-	mem := d.getEffectiveMem()
-	if isMemOverlaps(d.s.icmp, mem, r.Start, r.Limit) {
+	mem := db.getEffectiveMem()
+	defer mem.decref()
+	if isMemOverlaps(db.s.icmp, mem.mdb, r.Start, r.Limit) {
 		// Memdb compaction.
-		if _, err := d.rotateMem(0); err != nil {
-			<-d.writeLockC
+		if _, err := db.rotateMem(0); err != nil {
+			<-db.writeLockC
 			return err
 		}
-		<-d.writeLockC
-		if err := d.compSendIdle(d.mcompCmdC); err != nil {
+		<-db.writeLockC
+		if err := db.compSendIdle(db.mcompCmdC); err != nil {
 			return err
 		}
 	} else {
-		<-d.writeLockC
+		<-db.writeLockC
 	}
 
 	// Table compaction.
-	return d.compSendRange(d.tcompCmdC, -1, r.Start, r.Limit)
+	return db.compSendRange(db.tcompCmdC, -1, r.Start, r.Limit)
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/doc.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/doc.go
index ac9ea3d0c..53f13bb24 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/doc.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/doc.go
@@ -37,6 +37,16 @@
 //	err = iter.Error()
 //	...
 //
+// Iterate over subset of database content with a particular prefix:
+//	iter := db.NewIterator(util.BytesPrefix([]byte("foo-")), nil)
+//	for iter.Next() {
+//		// Use key/value.
+//		...
+//	}
+//	iter.Release()
+//	err = iter.Error()
+//	...
+//
 // Seek-then-Iterate:
 //
 // 	iter := db.NewIterator(nil, nil)
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/error.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/errors.go
index 8066bd9a4..29d0d2f27 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/error.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/errors.go
@@ -7,32 +7,12 @@
 package leveldb
 
 import (
-	"errors"
-
-	"github.com/syndtr/goleveldb/leveldb/util"
+	"github.com/syndtr/goleveldb/leveldb/errors"
 )
 
 var (
-	ErrNotFound         = util.ErrNotFound
+	ErrNotFound         = errors.ErrNotFound
 	ErrSnapshotReleased = errors.New("leveldb: snapshot released")
 	ErrIterReleased     = errors.New("leveldb: iterator released")
 	ErrClosed           = errors.New("leveldb: closed")
 )
-
-type CorruptionType int
-
-const (
-	CorruptedManifest CorruptionType = iota
-	MissingFiles
-)
-
-// ErrCorrupted is the type that wraps errors that indicate corruption in
-// the database.
-type ErrCorrupted struct {
-	Type CorruptionType
-	Err  error
-}
-
-func (e ErrCorrupted) Error() string {
-	return e.Err.Error()
-}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/errors/errors.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/errors/errors.go
new file mode 100644
index 000000000..84b5d6b7b
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/errors/errors.go
@@ -0,0 +1,76 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// Package errors provides common error types used throughout leveldb.
+package errors
+
+import (
+	"errors"
+	"fmt"
+
+	"github.com/syndtr/goleveldb/leveldb/storage"
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+var (
+	ErrNotFound    = New("leveldb: not found")
+	ErrReleased    = util.ErrReleased
+	ErrHasReleaser = util.ErrHasReleaser
+)
+
+// New returns an error that formats as the given text.
+func New(text string) error {
+	return errors.New(text)
+}
+
+// ErrCorrupted is the type that wraps errors that indicate corruption in
+// the database.
+type ErrCorrupted struct {
+	File *storage.FileInfo
+	Err  error
+}
+
+func (e *ErrCorrupted) Error() string {
+	if e.File != nil {
+		return fmt.Sprintf("%v [file=%v]", e.Err, e.File)
+	} else {
+		return e.Err.Error()
+	}
+}
+
+// NewErrCorrupted creates new ErrCorrupted error.
+func NewErrCorrupted(f storage.File, err error) error {
+	return &ErrCorrupted{storage.NewFileInfo(f), err}
+}
+
+// IsCorrupted returns a boolean indicating whether the error is indicating
+// a corruption.
+func IsCorrupted(err error) bool {
+	switch err.(type) {
+	case *ErrCorrupted:
+		return true
+	}
+	return false
+}
+
+// ErrMissingFiles is the type that indicating a corruption due to missing
+// files.
+type ErrMissingFiles struct {
+	Files []*storage.FileInfo
+}
+
+func (e *ErrMissingFiles) Error() string { return "file missing" }
+
+// SetFile sets 'file info' of the given error with the given file.
+// Currently only ErrCorrupted is supported, otherwise will do nothing.
+func SetFile(err error, f storage.File) error {
+	switch x := err.(type) {
+	case *ErrCorrupted:
+		x.File = storage.NewFileInfo(f)
+		return x
+	}
+	return err
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/external_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/external_test.go
index d7dff04b6..b328ece4e 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/external_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/external_test.go
@@ -17,13 +17,14 @@ import (
 var _ = testutil.Defer(func() {
 	Describe("Leveldb external", func() {
 		o := &opt.Options{
-			BlockCache:           opt.NoCache,
-			BlockRestartInterval: 5,
-			BlockSize:            50,
-			Compression:          opt.NoCompression,
-			MaxOpenFiles:         0,
-			Strict:               opt.StrictAll,
-			WriteBuffer:          1000,
+			DisableBlockCache:      true,
+			BlockRestartInterval:   5,
+			BlockSize:              80,
+			Compression:            opt.NoCompression,
+			OpenFilesCacheCapacity: -1,
+			Strict:                 opt.StrictAll,
+			WriteBuffer:            1000,
+			CompactionTableSize:    2000,
 		}
 
 		Describe("write test", func() {
@@ -36,22 +37,21 @@ var _ = testutil.Defer(func() {
 				testutil.DoDBTesting(&t)
 				db.TestClose()
 				done <- true
-			}, 9.0)
+			}, 20.0)
 		})
 
 		Describe("read test", func() {
-			testutil.AllKeyValueTesting(nil, func(kv testutil.KeyValue) testutil.DB {
+			testutil.AllKeyValueTesting(nil, nil, func(kv testutil.KeyValue) testutil.DB {
 				// Building the DB.
 				db := newTestingDB(o, nil, nil)
 				kv.IterateShuffled(nil, func(i int, key, value []byte) {
 					err := db.TestPut(key, value)
 					Expect(err).NotTo(HaveOccurred())
 				})
-				testutil.Defer("teardown", func() {
-					db.TestClose()
-				})
 
 				return db
+			}, func(db testutil.DB) {
+				db.(*testingDB).TestClose()
 			})
 		})
 	})
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/array_iter.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/array_iter.go
index 9b4b72741..a23ab05f7 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/array_iter.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/array_iter.go
@@ -40,13 +40,19 @@ type basicArrayIterator struct {
 	util.BasicReleaser
 	array BasicArray
 	pos   int
+	err   error
 }
 
 func (i *basicArrayIterator) Valid() bool {
-	return i.pos >= 0 && i.pos < i.array.Len()
+	return i.pos >= 0 && i.pos < i.array.Len() && !i.Released()
 }
 
 func (i *basicArrayIterator) First() bool {
+	if i.Released() {
+		i.err = ErrIterReleased
+		return false
+	}
+
 	if i.array.Len() == 0 {
 		i.pos = -1
 		return false
@@ -56,6 +62,11 @@ func (i *basicArrayIterator) First() bool {
 }
 
 func (i *basicArrayIterator) Last() bool {
+	if i.Released() {
+		i.err = ErrIterReleased
+		return false
+	}
+
 	n := i.array.Len()
 	if n == 0 {
 		i.pos = 0
@@ -66,6 +77,11 @@ func (i *basicArrayIterator) Last() bool {
 }
 
 func (i *basicArrayIterator) Seek(key []byte) bool {
+	if i.Released() {
+		i.err = ErrIterReleased
+		return false
+	}
+
 	n := i.array.Len()
 	if n == 0 {
 		i.pos = 0
@@ -79,6 +95,11 @@ func (i *basicArrayIterator) Seek(key []byte) bool {
 }
 
 func (i *basicArrayIterator) Next() bool {
+	if i.Released() {
+		i.err = ErrIterReleased
+		return false
+	}
+
 	i.pos++
 	if n := i.array.Len(); i.pos >= n {
 		i.pos = n
@@ -88,6 +109,11 @@ func (i *basicArrayIterator) Next() bool {
 }
 
 func (i *basicArrayIterator) Prev() bool {
+	if i.Released() {
+		i.err = ErrIterReleased
+		return false
+	}
+
 	i.pos--
 	if i.pos < 0 {
 		i.pos = -1
@@ -96,7 +122,7 @@ func (i *basicArrayIterator) Prev() bool {
 	return true
 }
 
-func (i *basicArrayIterator) Error() error { return nil }
+func (i *basicArrayIterator) Error() error { return i.err }
 
 type arrayIterator struct {
 	basicArrayIterator
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter.go
index 1e99a2bf6..939adbb93 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter.go
@@ -7,6 +7,7 @@
 package iterator
 
 import (
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
@@ -22,13 +23,13 @@ type IteratorIndexer interface {
 
 type indexedIterator struct {
 	util.BasicReleaser
-	index     IteratorIndexer
-	strict    bool
-	strictGet bool
+	index  IteratorIndexer
+	strict bool
 
-	data Iterator
-	err  error
-	errf func(err error)
+	data   Iterator
+	err    error
+	errf   func(err error)
+	closed bool
 }
 
 func (i *indexedIterator) setData() {
@@ -36,11 +37,6 @@ func (i *indexedIterator) setData() {
 		i.data.Release()
 	}
 	i.data = i.index.Get()
-	if i.strictGet {
-		if err := i.data.Error(); err != nil {
-			i.err = err
-		}
-	}
 }
 
 func (i *indexedIterator) clearData() {
@@ -50,14 +46,21 @@ func (i *indexedIterator) clearData() {
 	i.data = nil
 }
 
-func (i *indexedIterator) dataErr() bool {
-	if i.errf != nil {
-		if err := i.data.Error(); err != nil {
+func (i *indexedIterator) indexErr() {
+	if err := i.index.Error(); err != nil {
+		if i.errf != nil {
 			i.errf(err)
 		}
+		i.err = err
 	}
-	if i.strict {
-		if err := i.data.Error(); err != nil {
+}
+
+func (i *indexedIterator) dataErr() bool {
+	if err := i.data.Error(); err != nil {
+		if i.errf != nil {
+			i.errf(err)
+		}
+		if i.strict || !errors.IsCorrupted(err) {
 			i.err = err
 			return true
 		}
@@ -72,9 +75,13 @@ func (i *indexedIterator) Valid() bool {
 func (i *indexedIterator) First() bool {
 	if i.err != nil {
 		return false
+	} else if i.Released() {
+		i.err = ErrIterReleased
+		return false
 	}
 
 	if !i.index.First() {
+		i.indexErr()
 		i.clearData()
 		return false
 	}
@@ -85,9 +92,13 @@ func (i *indexedIterator) First() bool {
 func (i *indexedIterator) Last() bool {
 	if i.err != nil {
 		return false
+	} else if i.Released() {
+		i.err = ErrIterReleased
+		return false
 	}
 
 	if !i.index.Last() {
+		i.indexErr()
 		i.clearData()
 		return false
 	}
@@ -105,9 +116,13 @@ func (i *indexedIterator) Last() bool {
 func (i *indexedIterator) Seek(key []byte) bool {
 	if i.err != nil {
 		return false
+	} else if i.Released() {
+		i.err = ErrIterReleased
+		return false
 	}
 
 	if !i.index.Seek(key) {
+		i.indexErr()
 		i.clearData()
 		return false
 	}
@@ -125,6 +140,9 @@ func (i *indexedIterator) Seek(key []byte) bool {
 func (i *indexedIterator) Next() bool {
 	if i.err != nil {
 		return false
+	} else if i.Released() {
+		i.err = ErrIterReleased
+		return false
 	}
 
 	switch {
@@ -136,6 +154,7 @@ func (i *indexedIterator) Next() bool {
 		fallthrough
 	case i.data == nil:
 		if !i.index.Next() {
+			i.indexErr()
 			return false
 		}
 		i.setData()
@@ -147,6 +166,9 @@ func (i *indexedIterator) Next() bool {
 func (i *indexedIterator) Prev() bool {
 	if i.err != nil {
 		return false
+	} else if i.Released() {
+		i.err = ErrIterReleased
+		return false
 	}
 
 	switch {
@@ -158,6 +180,7 @@ func (i *indexedIterator) Prev() bool {
 		fallthrough
 	case i.data == nil:
 		if !i.index.Prev() {
+			i.indexErr()
 			return false
 		}
 		i.setData()
@@ -206,16 +229,14 @@ func (i *indexedIterator) SetErrorCallback(f func(err error)) {
 	i.errf = f
 }
 
-// NewIndexedIterator returns an indexed iterator. An index is iterator
-// that returns another iterator, a data iterator. A data iterator is the
+// NewIndexedIterator returns an 'indexed iterator'. An index is iterator
+// that returns another iterator, a 'data iterator'. A 'data iterator' is the
 // iterator that contains actual key/value pairs.
 //
-// If strict is true then error yield by data iterator will halt the indexed
-// iterator, on contrary if strict is false then the indexed iterator will
-// ignore those error and move on to the next index. If strictGet is true and
-// index.Get() yield an 'error iterator' then the indexed iterator will be halted.
-// An 'error iterator' is iterator which its Error() method always return non-nil
-// even before any 'seeks method' is called.
-func NewIndexedIterator(index IteratorIndexer, strict, strictGet bool) Iterator {
-	return &indexedIterator{index: index, strict: strict, strictGet: strictGet}
+// If strict is true the any 'corruption errors' (i.e errors.IsCorrupted(err) == true)
+// won't be ignored and will halt 'indexed iterator', otherwise the iterator will
+// continue to the next 'data iterator'. Corruption on 'index iterator' will not be
+// ignored and will halt the iterator.
+func NewIndexedIterator(index IteratorIndexer, strict bool) Iterator {
+	return &indexedIterator{index: index, strict: strict}
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter_test.go
index 6a89b3830..72a797892 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/indexed_iter_test.go
@@ -65,7 +65,7 @@ var _ = testutil.Defer(func() {
 					// Test the iterator.
 					t := testutil.IteratorTesting{
 						KeyValue: kv.Clone(),
-						Iter:     NewIndexedIterator(NewArrayIndexer(index), true, true),
+						Iter:     NewIndexedIterator(NewArrayIndexer(index), true),
 					}
 					testutil.DoIteratorTesting(&t)
 					done <- true
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter.go
index 1b80184e8..c2522860b 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter.go
@@ -14,6 +14,10 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
+var (
+	ErrIterReleased = errors.New("leveldb/iterator: iterator released")
+)
+
 // IteratorSeeker is the interface that wraps the 'seeks method'.
 type IteratorSeeker interface {
 	// First moves the iterator to the first key/value pair. If the iterator
@@ -100,28 +104,13 @@ type ErrorCallbackSetter interface {
 }
 
 type emptyIterator struct {
-	releaser util.Releaser
-	released bool
-	err      error
+	util.BasicReleaser
+	err error
 }
 
 func (i *emptyIterator) rErr() {
-	if i.err == nil && i.released {
-		i.err = errors.New("leveldb/iterator: iterator released")
-	}
-}
-
-func (i *emptyIterator) Release() {
-	if i.releaser != nil {
-		i.releaser.Release()
-		i.releaser = nil
-	}
-	i.released = true
-}
-
-func (i *emptyIterator) SetReleaser(releaser util.Releaser) {
-	if !i.released {
-		i.releaser = releaser
+	if i.err == nil && i.Released() {
+		i.err = ErrIterReleased
 	}
 }
 
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter_suite_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter_suite_test.go
index 7ec2fc6f2..5ef8d5baf 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter_suite_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/iter_suite_test.go
@@ -3,15 +3,9 @@ package iterator_test
 import (
 	"testing"
 
-	. "github.com/onsi/ginkgo"
-	. "github.com/onsi/gomega"
-
 	"github.com/syndtr/goleveldb/leveldb/testutil"
 )
 
 func TestIterator(t *testing.T) {
-	testutil.RunDefer()
-
-	RegisterFailHandler(Fail)
-	RunSpecs(t, "Iterator Suite")
+	testutil.RunSuite(t, "Iterator Suite")
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/merged_iter.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/merged_iter.go
index c8314c4e5..1a7e29df8 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/merged_iter.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/iterator/merged_iter.go
@@ -7,16 +7,11 @@
 package iterator
 
 import (
-	"errors"
-
 	"github.com/syndtr/goleveldb/leveldb/comparer"
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-var (
-	ErrIterReleased = errors.New("leveldb/iterator: iterator released")
-)
-
 type dir int
 
 const (
@@ -48,13 +43,11 @@ func assertKey(key []byte) []byte {
 }
 
 func (i *mergedIterator) iterErr(iter Iterator) bool {
-	if i.errf != nil {
-		if err := iter.Error(); err != nil {
+	if err := iter.Error(); err != nil {
+		if i.errf != nil {
 			i.errf(err)
 		}
-	}
-	if i.strict {
-		if err := iter.Error(); err != nil {
+		if i.strict || !errors.IsCorrupted(err) {
 			i.err = err
 			return true
 		}
@@ -274,9 +267,13 @@ func (i *mergedIterator) Release() {
 }
 
 func (i *mergedIterator) SetReleaser(releaser util.Releaser) {
-	if i.dir != dirReleased {
-		i.releaser = releaser
+	if i.dir == dirReleased {
+		panic(util.ErrReleased)
+	}
+	if i.releaser != nil && releaser != nil {
+		panic(util.ErrHasReleaser)
 	}
+	i.releaser = releaser
 }
 
 func (i *mergedIterator) Error() error {
@@ -294,9 +291,9 @@ func (i *mergedIterator) SetErrorCallback(f func(err error)) {
 // keys: if iters[i] contains a key k then iters[j] will not contain that key k.
 // None of the iters may be nil.
 //
-// If strict is true then error yield by any iterators will halt the merged
-// iterator, on contrary if strict is false then the merged iterator will
-// ignore those error and move on to the next iterator.
+// If strict is true the any 'corruption errors' (i.e errors.IsCorrupted(err) == true)
+// won't be ignored and will halt 'merged iterator', otherwise the iterator will
+// continue to the next 'input iterator'.
 func NewMergedIterator(iters []Iterator, cmp comparer.Comparer, strict bool) Iterator {
 	return &mergedIterator{
 		iters:  iters,
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal.go
index b522c76e6..6519ec660 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal.go
@@ -79,10 +79,10 @@ package journal
 
 import (
 	"encoding/binary"
-	"errors"
 	"fmt"
 	"io"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
@@ -103,18 +103,18 @@ type flusher interface {
 	Flush() error
 }
 
-// DroppedError is the error type that passed to Dropper.Drop method.
-type DroppedError struct {
+// ErrCorrupted is the error type that generated by corrupted block or chunk.
+type ErrCorrupted struct {
 	Size   int
 	Reason string
 }
 
-func (e DroppedError) Error() string {
-	return fmt.Sprintf("leveldb/journal: dropped %d bytes: %s", e.Size, e.Reason)
+func (e *ErrCorrupted) Error() string {
+	return fmt.Sprintf("leveldb/journal: block/chunk corrupted: %s (%d bytes)", e.Reason, e.Size)
 }
 
 // Dropper is the interface that wrap simple Drop method. The Drop
-// method will be called when the journal reader dropping a chunk.
+// method will be called when the journal reader dropping a block or chunk.
 type Dropper interface {
 	Drop(err error)
 }
@@ -158,76 +158,78 @@ func NewReader(r io.Reader, dropper Dropper, strict, checksum bool) *Reader {
 	}
 }
 
+var errSkip = errors.New("leveldb/journal: skipped")
+
+func (r *Reader) corrupt(n int, reason string, skip bool) error {
+	if r.dropper != nil {
+		r.dropper.Drop(&ErrCorrupted{n, reason})
+	}
+	if r.strict && !skip {
+		r.err = errors.NewErrCorrupted(nil, &ErrCorrupted{n, reason})
+		return r.err
+	}
+	return errSkip
+}
+
 // nextChunk sets r.buf[r.i:r.j] to hold the next chunk's payload, reading the
 // next block into the buffer if necessary.
-func (r *Reader) nextChunk(wantFirst, skip bool) error {
+func (r *Reader) nextChunk(first bool) error {
 	for {
 		if r.j+headerSize <= r.n {
 			checksum := binary.LittleEndian.Uint32(r.buf[r.j+0 : r.j+4])
 			length := binary.LittleEndian.Uint16(r.buf[r.j+4 : r.j+6])
 			chunkType := r.buf[r.j+6]
 
-			var err error
 			if checksum == 0 && length == 0 && chunkType == 0 {
 				// Drop entire block.
-				err = DroppedError{r.n - r.j, "zero header"}
+				m := r.n - r.j
 				r.i = r.n
 				r.j = r.n
+				return r.corrupt(m, "zero header", false)
 			} else {
 				m := r.n - r.j
 				r.i = r.j + headerSize
 				r.j = r.j + headerSize + int(length)
 				if r.j > r.n {
 					// Drop entire block.
-					err = DroppedError{m, "chunk length overflows block"}
 					r.i = r.n
 					r.j = r.n
+					return r.corrupt(m, "chunk length overflows block", false)
 				} else if r.checksum && checksum != util.NewCRC(r.buf[r.i-1:r.j]).Value() {
 					// Drop entire block.
-					err = DroppedError{m, "checksum mismatch"}
 					r.i = r.n
 					r.j = r.n
+					return r.corrupt(m, "checksum mismatch", false)
 				}
 			}
-			if wantFirst && err == nil && chunkType != fullChunkType && chunkType != firstChunkType {
-				if skip {
-					// The chunk are intentionally skipped.
-					if chunkType == lastChunkType {
-						skip = false
-					}
-					continue
-				} else {
-					// Drop the chunk.
-					err = DroppedError{r.j - r.i + headerSize, "orphan chunk"}
-				}
-			}
-			if err == nil {
-				r.last = chunkType == fullChunkType || chunkType == lastChunkType
-			} else {
-				if r.dropper != nil {
-					r.dropper.Drop(err)
-				}
-				if r.strict {
-					r.err = err
-				}
+			if first && chunkType != fullChunkType && chunkType != firstChunkType {
+				m := r.j - r.i
+				r.i = r.j
+				// Report the error, but skip it.
+				return r.corrupt(m+headerSize, "orphan chunk", true)
 			}
-			return err
+			r.last = chunkType == fullChunkType || chunkType == lastChunkType
+			return nil
 		}
+
+		// The last block.
 		if r.n < blockSize && r.n > 0 {
-			// This is the last block.
-			if r.j != r.n {
-				r.err = io.ErrUnexpectedEOF
-			} else {
-				r.err = io.EOF
+			if !first {
+				return r.corrupt(0, "missing chunk part", false)
 			}
+			r.err = io.EOF
 			return r.err
 		}
+
+		// Read block.
 		n, err := io.ReadFull(r.r, r.buf[:])
-		if err != nil && err != io.ErrUnexpectedEOF {
-			r.err = err
-			return r.err
+		if err != nil && err != io.EOF && err != io.ErrUnexpectedEOF {
+			return err
 		}
 		if n == 0 {
+			if !first {
+				return r.corrupt(0, "missing chunk part", false)
+			}
 			r.err = io.EOF
 			return r.err
 		}
@@ -237,29 +239,26 @@ func (r *Reader) nextChunk(wantFirst, skip bool) error {
 
 // Next returns a reader for the next journal. It returns io.EOF if there are no
 // more journals. The reader returned becomes stale after the next Next call,
-// and should no longer be used.
+// and should no longer be used. If strict is false, the reader will returns
+// io.ErrUnexpectedEOF error when found corrupted journal.
 func (r *Reader) Next() (io.Reader, error) {
 	r.seq++
 	if r.err != nil {
 		return nil, r.err
 	}
-	skip := !r.last
+	r.i = r.j
 	for {
-		r.i = r.j
-		if r.nextChunk(true, skip) != nil {
-			// So that 'orphan chunk' drop will be reported.
-			skip = false
-		} else {
+		if err := r.nextChunk(true); err == nil {
 			break
-		}
-		if r.err != nil {
-			return nil, r.err
+		} else if err != errSkip {
+			return nil, err
 		}
 	}
 	return &singleReader{r, r.seq, nil}, nil
 }
 
-// Reset resets the journal reader, allows reuse of the journal reader.
+// Reset resets the journal reader, allows reuse of the journal reader. Reset returns
+// last accumulated error.
 func (r *Reader) Reset(reader io.Reader, dropper Dropper, strict, checksum bool) error {
 	r.seq++
 	err := r.err
@@ -296,7 +295,11 @@ func (x *singleReader) Read(p []byte) (int, error) {
 		if r.last {
 			return 0, io.EOF
 		}
-		if x.err = r.nextChunk(false, false); x.err != nil {
+		x.err = r.nextChunk(false)
+		if x.err != nil {
+			if x.err == errSkip {
+				x.err = io.ErrUnexpectedEOF
+			}
 			return 0, x.err
 		}
 	}
@@ -320,7 +323,11 @@ func (x *singleReader) ReadByte() (byte, error) {
 		if r.last {
 			return 0, io.EOF
 		}
-		if x.err = r.nextChunk(false, false); x.err != nil {
+		x.err = r.nextChunk(false)
+		if x.err != nil {
+			if x.err == errSkip {
+				x.err = io.ErrUnexpectedEOF
+			}
 			return 0, x.err
 		}
 	}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal_test.go
index 5e1193ae2..0fcf22599 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/journal/journal_test.go
@@ -12,6 +12,7 @@ package journal
 
 import (
 	"bytes"
+	"encoding/binary"
 	"fmt"
 	"io"
 	"io/ioutil"
@@ -326,3 +327,492 @@ func TestStaleWriter(t *testing.T) {
 		t.Fatalf("stale write #1: unexpected error: %v", err)
 	}
 }
+
+func TestCorrupt_MissingLastBlock(t *testing.T) {
+	buf := new(bytes.Buffer)
+
+	w := NewWriter(buf)
+
+	// First record.
+	ww, err := w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-1024)); err != nil {
+		t.Fatalf("write #0: unexpected error: %v", err)
+	}
+
+	// Second record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-headerSize)); err != nil {
+		t.Fatalf("write #1: unexpected error: %v", err)
+	}
+
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	// Cut the last block.
+	b := buf.Bytes()[:blockSize]
+	r := NewReader(bytes.NewReader(b), dropper{t}, false, true)
+
+	// First read.
+	rr, err := r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err := io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #0: %v", err)
+	}
+	if n != blockSize-1024 {
+		t.Fatalf("read #0: got %d bytes want %d", n, blockSize-1024)
+	}
+
+	// Second read.
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != io.ErrUnexpectedEOF {
+		t.Fatalf("read #1: unexpected error: %v", err)
+	}
+
+	if _, err := r.Next(); err != io.EOF {
+		t.Fatalf("last next: unexpected error: %v", err)
+	}
+}
+
+func TestCorrupt_CorruptedFirstBlock(t *testing.T) {
+	buf := new(bytes.Buffer)
+
+	w := NewWriter(buf)
+
+	// First record.
+	ww, err := w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize/2)); err != nil {
+		t.Fatalf("write #0: unexpected error: %v", err)
+	}
+
+	// Second record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-headerSize)); err != nil {
+		t.Fatalf("write #1: unexpected error: %v", err)
+	}
+
+	// Third record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+1)); err != nil {
+		t.Fatalf("write #2: unexpected error: %v", err)
+	}
+
+	// Fourth record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+2)); err != nil {
+		t.Fatalf("write #3: unexpected error: %v", err)
+	}
+
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	b := buf.Bytes()
+	// Corrupting block #0.
+	for i := 0; i < 1024; i++ {
+		b[i] = '1'
+	}
+
+	r := NewReader(bytes.NewReader(b), dropper{t}, false, true)
+
+	// First read (third record).
+	rr, err := r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err := io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #0: %v", err)
+	}
+	if want := int64(blockSize-headerSize) + 1; n != want {
+		t.Fatalf("read #0: got %d bytes want %d", n, want)
+	}
+
+	// Second read (fourth record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #1: %v", err)
+	}
+	if want := int64(blockSize-headerSize) + 2; n != want {
+		t.Fatalf("read #1: got %d bytes want %d", n, want)
+	}
+
+	if _, err := r.Next(); err != io.EOF {
+		t.Fatalf("last next: unexpected error: %v", err)
+	}
+}
+
+func TestCorrupt_CorruptedMiddleBlock(t *testing.T) {
+	buf := new(bytes.Buffer)
+
+	w := NewWriter(buf)
+
+	// First record.
+	ww, err := w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize/2)); err != nil {
+		t.Fatalf("write #0: unexpected error: %v", err)
+	}
+
+	// Second record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-headerSize)); err != nil {
+		t.Fatalf("write #1: unexpected error: %v", err)
+	}
+
+	// Third record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+1)); err != nil {
+		t.Fatalf("write #2: unexpected error: %v", err)
+	}
+
+	// Fourth record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+2)); err != nil {
+		t.Fatalf("write #3: unexpected error: %v", err)
+	}
+
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	b := buf.Bytes()
+	// Corrupting block #1.
+	for i := 0; i < 1024; i++ {
+		b[blockSize+i] = '1'
+	}
+
+	r := NewReader(bytes.NewReader(b), dropper{t}, false, true)
+
+	// First read (first record).
+	rr, err := r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err := io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #0: %v", err)
+	}
+	if want := int64(blockSize / 2); n != want {
+		t.Fatalf("read #0: got %d bytes want %d", n, want)
+	}
+
+	// Second read (second record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != io.ErrUnexpectedEOF {
+		t.Fatalf("read #1: unexpected error: %v", err)
+	}
+
+	// Third read (fourth record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #2: %v", err)
+	}
+	if want := int64(blockSize-headerSize) + 2; n != want {
+		t.Fatalf("read #2: got %d bytes want %d", n, want)
+	}
+
+	if _, err := r.Next(); err != io.EOF {
+		t.Fatalf("last next: unexpected error: %v", err)
+	}
+}
+
+func TestCorrupt_CorruptedLastBlock(t *testing.T) {
+	buf := new(bytes.Buffer)
+
+	w := NewWriter(buf)
+
+	// First record.
+	ww, err := w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize/2)); err != nil {
+		t.Fatalf("write #0: unexpected error: %v", err)
+	}
+
+	// Second record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-headerSize)); err != nil {
+		t.Fatalf("write #1: unexpected error: %v", err)
+	}
+
+	// Third record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+1)); err != nil {
+		t.Fatalf("write #2: unexpected error: %v", err)
+	}
+
+	// Fourth record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+2)); err != nil {
+		t.Fatalf("write #3: unexpected error: %v", err)
+	}
+
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	b := buf.Bytes()
+	// Corrupting block #3.
+	for i := len(b) - 1; i > len(b)-1024; i-- {
+		b[i] = '1'
+	}
+
+	r := NewReader(bytes.NewReader(b), dropper{t}, false, true)
+
+	// First read (first record).
+	rr, err := r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err := io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #0: %v", err)
+	}
+	if want := int64(blockSize / 2); n != want {
+		t.Fatalf("read #0: got %d bytes want %d", n, want)
+	}
+
+	// Second read (second record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #1: %v", err)
+	}
+	if want := int64(blockSize - headerSize); n != want {
+		t.Fatalf("read #1: got %d bytes want %d", n, want)
+	}
+
+	// Third read (third record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #2: %v", err)
+	}
+	if want := int64(blockSize-headerSize) + 1; n != want {
+		t.Fatalf("read #2: got %d bytes want %d", n, want)
+	}
+
+	// Fourth read (fourth record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != io.ErrUnexpectedEOF {
+		t.Fatalf("read #3: unexpected error: %v", err)
+	}
+
+	if _, err := r.Next(); err != io.EOF {
+		t.Fatalf("last next: unexpected error: %v", err)
+	}
+}
+
+func TestCorrupt_FirstChuckLengthOverflow(t *testing.T) {
+	buf := new(bytes.Buffer)
+
+	w := NewWriter(buf)
+
+	// First record.
+	ww, err := w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize/2)); err != nil {
+		t.Fatalf("write #0: unexpected error: %v", err)
+	}
+
+	// Second record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-headerSize)); err != nil {
+		t.Fatalf("write #1: unexpected error: %v", err)
+	}
+
+	// Third record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+1)); err != nil {
+		t.Fatalf("write #2: unexpected error: %v", err)
+	}
+
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	b := buf.Bytes()
+	// Corrupting record #1.
+	x := blockSize
+	binary.LittleEndian.PutUint16(b[x+4:], 0xffff)
+
+	r := NewReader(bytes.NewReader(b), dropper{t}, false, true)
+
+	// First read (first record).
+	rr, err := r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err := io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #0: %v", err)
+	}
+	if want := int64(blockSize / 2); n != want {
+		t.Fatalf("read #0: got %d bytes want %d", n, want)
+	}
+
+	// Second read (second record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != io.ErrUnexpectedEOF {
+		t.Fatalf("read #1: unexpected error: %v", err)
+	}
+
+	if _, err := r.Next(); err != io.EOF {
+		t.Fatalf("last next: unexpected error: %v", err)
+	}
+}
+
+func TestCorrupt_MiddleChuckLengthOverflow(t *testing.T) {
+	buf := new(bytes.Buffer)
+
+	w := NewWriter(buf)
+
+	// First record.
+	ww, err := w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize/2)); err != nil {
+		t.Fatalf("write #0: unexpected error: %v", err)
+	}
+
+	// Second record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), blockSize-headerSize)); err != nil {
+		t.Fatalf("write #1: unexpected error: %v", err)
+	}
+
+	// Third record.
+	ww, err = w.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := ww.Write(bytes.Repeat([]byte("0"), (blockSize-headerSize)+1)); err != nil {
+		t.Fatalf("write #2: unexpected error: %v", err)
+	}
+
+	if err := w.Close(); err != nil {
+		t.Fatal(err)
+	}
+
+	b := buf.Bytes()
+	// Corrupting record #1.
+	x := blockSize/2 + headerSize
+	binary.LittleEndian.PutUint16(b[x+4:], 0xffff)
+
+	r := NewReader(bytes.NewReader(b), dropper{t}, false, true)
+
+	// First read (first record).
+	rr, err := r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err := io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #0: %v", err)
+	}
+	if want := int64(blockSize / 2); n != want {
+		t.Fatalf("read #0: got %d bytes want %d", n, want)
+	}
+
+	// Second read (third record).
+	rr, err = r.Next()
+	if err != nil {
+		t.Fatal(err)
+	}
+	n, err = io.Copy(ioutil.Discard, rr)
+	if err != nil {
+		t.Fatalf("read #1: %v", err)
+	}
+	if want := int64(blockSize-headerSize) + 1; n != want {
+		t.Fatalf("read #1: got %d bytes want %d", n, want)
+	}
+
+	if _, err := r.Next(); err != io.EOF {
+		t.Fatalf("last next: unexpected error: %v", err)
+	}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key.go
index b9acf932d..572ae8150 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key.go
@@ -9,15 +9,30 @@ package leveldb
 import (
 	"encoding/binary"
 	"fmt"
+
+	"github.com/syndtr/goleveldb/leveldb/errors"
 )
 
-type vType int
+type ErrIkeyCorrupted struct {
+	Ikey   []byte
+	Reason string
+}
+
+func (e *ErrIkeyCorrupted) Error() string {
+	return fmt.Sprintf("leveldb: iKey %q corrupted: %s", e.Ikey, e.Reason)
+}
+
+func newErrIkeyCorrupted(ikey []byte, reason string) error {
+	return errors.NewErrCorrupted(nil, &ErrIkeyCorrupted{append([]byte{}, ikey...), reason})
+}
+
+type kType int
 
-func (t vType) String() string {
-	switch t {
-	case tDel:
+func (kt kType) String() string {
+	switch kt {
+	case ktDel:
 		return "d"
-	case tVal:
+	case ktVal:
 		return "v"
 	}
 	return "x"
@@ -26,16 +41,16 @@ func (t vType) String() string {
 // Value types encoded as the last component of internal keys.
 // Don't modify; this value are saved to disk.
 const (
-	tDel vType = iota
-	tVal
+	ktDel kType = iota
+	ktVal
 )
 
-// tSeek defines the vType that should be passed when constructing an
+// ktSeek defines the kType that should be passed when constructing an
 // internal key for seeking to a particular sequence number (since we
 // sort sequence numbers in decreasing order and the value type is
 // embedded as the low 8 bits in the sequence number in internal keys,
 // we need to use the highest-numbered ValueType, not the lowest).
-const tSeek = tVal
+const ktSeek = ktVal
 
 const (
 	// Maximum value possible for sequence number; the 8-bits are
@@ -43,7 +58,7 @@ const (
 	// 64-bit integer.
 	kMaxSeq uint64 = (uint64(1) << 56) - 1
 	// Maximum value possible for packed sequence number and type.
-	kMaxNum uint64 = (kMaxSeq << 8) | uint64(tSeek)
+	kMaxNum uint64 = (kMaxSeq << 8) | uint64(ktSeek)
 )
 
 // Maximum number encoded in bytes.
@@ -55,85 +70,73 @@ func init() {
 
 type iKey []byte
 
-func newIKey(ukey []byte, seq uint64, t vType) iKey {
-	if seq > kMaxSeq || t > tVal {
-		panic("invalid seq number or value type")
+func newIkey(ukey []byte, seq uint64, kt kType) iKey {
+	if seq > kMaxSeq {
+		panic("leveldb: invalid sequence number")
+	} else if kt > ktVal {
+		panic("leveldb: invalid type")
 	}
 
-	b := make(iKey, len(ukey)+8)
-	copy(b, ukey)
-	binary.LittleEndian.PutUint64(b[len(ukey):], (seq<<8)|uint64(t))
-	return b
+	ik := make(iKey, len(ukey)+8)
+	copy(ik, ukey)
+	binary.LittleEndian.PutUint64(ik[len(ukey):], (seq<<8)|uint64(kt))
+	return ik
 }
 
-func parseIkey(p []byte) (ukey []byte, seq uint64, t vType, ok bool) {
-	if len(p) < 8 {
-		return
+func parseIkey(ik []byte) (ukey []byte, seq uint64, kt kType, err error) {
+	if len(ik) < 8 {
+		return nil, 0, 0, newErrIkeyCorrupted(ik, "invalid length")
 	}
-	num := binary.LittleEndian.Uint64(p[len(p)-8:])
-	seq, t = uint64(num>>8), vType(num&0xff)
-	if t > tVal {
-		return
+	num := binary.LittleEndian.Uint64(ik[len(ik)-8:])
+	seq, kt = uint64(num>>8), kType(num&0xff)
+	if kt > ktVal {
+		return nil, 0, 0, newErrIkeyCorrupted(ik, "invalid type")
 	}
-	ukey = p[:len(p)-8]
-	ok = true
+	ukey = ik[:len(ik)-8]
 	return
 }
 
-func validIkey(p []byte) bool {
-	_, _, _, ok := parseIkey(p)
-	return ok
+func validIkey(ik []byte) bool {
+	_, _, _, err := parseIkey(ik)
+	return err == nil
 }
 
-func (p iKey) assert() {
-	if p == nil {
-		panic("nil iKey")
+func (ik iKey) assert() {
+	if ik == nil {
+		panic("leveldb: nil iKey")
 	}
-	if len(p) < 8 {
-		panic(fmt.Sprintf("invalid iKey %q, len=%d", []byte(p), len(p)))
+	if len(ik) < 8 {
+		panic(fmt.Sprintf("leveldb: iKey %q, len=%d: invalid length", []byte(ik), len(ik)))
 	}
 }
 
-func (p iKey) ok() bool {
-	if len(p) < 8 {
-		return false
-	}
-	_, _, ok := p.parseNum()
-	return ok
-}
-
-func (p iKey) ukey() []byte {
-	p.assert()
-	return p[:len(p)-8]
+func (ik iKey) ukey() []byte {
+	ik.assert()
+	return ik[:len(ik)-8]
 }
 
-func (p iKey) num() uint64 {
-	p.assert()
-	return binary.LittleEndian.Uint64(p[len(p)-8:])
+func (ik iKey) num() uint64 {
+	ik.assert()
+	return binary.LittleEndian.Uint64(ik[len(ik)-8:])
 }
 
-func (p iKey) parseNum() (seq uint64, t vType, ok bool) {
-	if p == nil {
-		panic("nil iKey")
+func (ik iKey) parseNum() (seq uint64, kt kType) {
+	num := ik.num()
+	seq, kt = uint64(num>>8), kType(num&0xff)
+	if kt > ktVal {
+		panic(fmt.Sprintf("leveldb: iKey %q, len=%d: invalid type %#x", []byte(ik), len(ik), kt))
 	}
-	if len(p) < 8 {
-		return
-	}
-	num := p.num()
-	seq, t = uint64(num>>8), vType(num&0xff)
-	if t > tVal {
-		return 0, 0, false
-	}
-	ok = true
 	return
 }
 
-func (p iKey) String() string {
-	if len(p) == 0 {
+func (ik iKey) String() string {
+	if ik == nil {
 		return "<nil>"
 	}
-	if seq, t, ok := p.parseNum(); ok {
-		return fmt.Sprintf("%s,%s%d", shorten(string(p.ukey())), t, seq)
+
+	if ukey, seq, kt, err := parseIkey(ik); err == nil {
+		return fmt.Sprintf("%s,%s%d", shorten(string(ukey)), kt, seq)
+	} else {
+		return "<invalid>"
 	}
-	return "<invalid>"
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key_test.go
index e307cfc1d..30eadf784 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/key_test.go
@@ -15,8 +15,8 @@ import (
 
 var defaultIComparer = &iComparer{comparer.DefaultComparer}
 
-func ikey(key string, seq uint64, t vType) iKey {
-	return newIKey([]byte(key), uint64(seq), t)
+func ikey(key string, seq uint64, kt kType) iKey {
+	return newIkey([]byte(key), uint64(seq), kt)
 }
 
 func shortSep(a, b []byte) []byte {
@@ -37,27 +37,37 @@ func shortSuccessor(b []byte) []byte {
 	return dst
 }
 
-func testSingleKey(t *testing.T, key string, seq uint64, vt vType) {
-	ik := ikey(key, seq, vt)
+func testSingleKey(t *testing.T, key string, seq uint64, kt kType) {
+	ik := ikey(key, seq, kt)
 
 	if !bytes.Equal(ik.ukey(), []byte(key)) {
 		t.Errorf("user key does not equal, got %v, want %v", string(ik.ukey()), key)
 	}
 
-	if rseq, rt, ok := ik.parseNum(); ok {
+	rseq, rt := ik.parseNum()
+	if rseq != seq {
+		t.Errorf("seq number does not equal, got %v, want %v", rseq, seq)
+	}
+	if rt != kt {
+		t.Errorf("type does not equal, got %v, want %v", rt, kt)
+	}
+
+	if rukey, rseq, rt, kerr := parseIkey(ik); kerr == nil {
+		if !bytes.Equal(rukey, []byte(key)) {
+			t.Errorf("user key does not equal, got %v, want %v", string(ik.ukey()), key)
+		}
 		if rseq != seq {
 			t.Errorf("seq number does not equal, got %v, want %v", rseq, seq)
 		}
-
-		if rt != vt {
-			t.Errorf("type does not equal, got %v, want %v", rt, vt)
+		if rt != kt {
+			t.Errorf("type does not equal, got %v, want %v", rt, kt)
 		}
 	} else {
-		t.Error("cannot parse seq and type")
+		t.Errorf("key error: %v", kerr)
 	}
 }
 
-func TestIKey_EncodeDecode(t *testing.T) {
+func TestIkey_EncodeDecode(t *testing.T) {
 	keys := []string{"", "k", "hello", "longggggggggggggggggggggg"}
 	seqs := []uint64{
 		1, 2, 3,
@@ -67,8 +77,8 @@ func TestIKey_EncodeDecode(t *testing.T) {
 	}
 	for _, key := range keys {
 		for _, seq := range seqs {
-			testSingleKey(t, key, seq, tVal)
-			testSingleKey(t, "hello", 1, tDel)
+			testSingleKey(t, key, seq, ktVal)
+			testSingleKey(t, "hello", 1, ktDel)
 		}
 	}
 }
@@ -79,45 +89,45 @@ func assertBytes(t *testing.T, want, got []byte) {
 	}
 }
 
-func TestIKeyShortSeparator(t *testing.T) {
+func TestIkeyShortSeparator(t *testing.T) {
 	// When user keys are same
-	assertBytes(t, ikey("foo", 100, tVal),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("foo", 99, tVal)))
-	assertBytes(t, ikey("foo", 100, tVal),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("foo", 101, tVal)))
-	assertBytes(t, ikey("foo", 100, tVal),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("foo", 100, tVal)))
-	assertBytes(t, ikey("foo", 100, tVal),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("foo", 100, tDel)))
+	assertBytes(t, ikey("foo", 100, ktVal),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("foo", 99, ktVal)))
+	assertBytes(t, ikey("foo", 100, ktVal),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("foo", 101, ktVal)))
+	assertBytes(t, ikey("foo", 100, ktVal),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("foo", 100, ktVal)))
+	assertBytes(t, ikey("foo", 100, ktVal),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("foo", 100, ktDel)))
 
 	// When user keys are misordered
-	assertBytes(t, ikey("foo", 100, tVal),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("bar", 99, tVal)))
+	assertBytes(t, ikey("foo", 100, ktVal),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("bar", 99, ktVal)))
 
 	// When user keys are different, but correctly ordered
-	assertBytes(t, ikey("g", uint64(kMaxSeq), tSeek),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("hello", 200, tVal)))
+	assertBytes(t, ikey("g", uint64(kMaxSeq), ktSeek),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("hello", 200, ktVal)))
 
 	// When start user key is prefix of limit user key
-	assertBytes(t, ikey("foo", 100, tVal),
-		shortSep(ikey("foo", 100, tVal),
-			ikey("foobar", 200, tVal)))
+	assertBytes(t, ikey("foo", 100, ktVal),
+		shortSep(ikey("foo", 100, ktVal),
+			ikey("foobar", 200, ktVal)))
 
 	// When limit user key is prefix of start user key
-	assertBytes(t, ikey("foobar", 100, tVal),
-		shortSep(ikey("foobar", 100, tVal),
-			ikey("foo", 200, tVal)))
+	assertBytes(t, ikey("foobar", 100, ktVal),
+		shortSep(ikey("foobar", 100, ktVal),
+			ikey("foo", 200, ktVal)))
 }
 
-func TestIKeyShortestSuccessor(t *testing.T) {
-	assertBytes(t, ikey("g", uint64(kMaxSeq), tSeek),
-		shortSuccessor(ikey("foo", 100, tVal)))
-	assertBytes(t, ikey("\xff\xff", 100, tVal),
-		shortSuccessor(ikey("\xff\xff", 100, tVal)))
+func TestIkeyShortestSuccessor(t *testing.T) {
+	assertBytes(t, ikey("g", uint64(kMaxSeq), ktSeek),
+		shortSuccessor(ikey("foo", 100, ktVal)))
+	assertBytes(t, ikey("\xff\xff", 100, ktVal),
+		shortSuccessor(ikey("\xff\xff", 100, ktVal)))
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/leveldb_suite_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/leveldb_suite_test.go
index 245b1fd4d..fefa007a7 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/leveldb_suite_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/leveldb_suite_test.go
@@ -3,18 +3,9 @@ package leveldb
 import (
 	"testing"
 
-	. "github.com/onsi/ginkgo"
-	. "github.com/onsi/gomega"
-
 	"github.com/syndtr/goleveldb/leveldb/testutil"
 )
 
-func TestLeveldb(t *testing.T) {
-	testutil.RunDefer()
-
-	RegisterFailHandler(Fail)
-	RunSpecs(t, "Leveldb Suite")
-
-	RegisterTestingT(t)
-	testutil.RunDefer("teardown")
+func TestLevelDB(t *testing.T) {
+	testutil.RunSuite(t, "LevelDB Suite")
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb.go
index 7bcae992a..e5398873b 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb.go
@@ -12,12 +12,14 @@ import (
 	"sync"
 
 	"github.com/syndtr/goleveldb/leveldb/comparer"
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
 var (
-	ErrNotFound = util.ErrNotFound
+	ErrNotFound     = errors.ErrNotFound
+	ErrIterReleased = errors.New("leveldb/memdb: iterator released")
 )
 
 const tMaxHeight = 12
@@ -29,6 +31,7 @@ type dbIter struct {
 	node       int
 	forward    bool
 	key, value []byte
+	err        error
 }
 
 func (i *dbIter) fill(checkStart, checkLimit bool) bool {
@@ -59,6 +62,11 @@ func (i *dbIter) Valid() bool {
 }
 
 func (i *dbIter) First() bool {
+	if i.Released() {
+		i.err = ErrIterReleased
+		return false
+	}
+
 	i.forward = true
 	i.p.mu.RLock()
 	defer i.p.mu.RUnlock()
@@ -71,9 +79,11 @@ func (i *dbIter) First() bool {
 }
 
 func (i *dbIter) Last() bool {
-	if i.p == nil {
+	if i.Released() {
+		i.err = ErrIterReleased
 		return false
 	}
+
 	i.forward = false
 	i.p.mu.RLock()
 	defer i.p.mu.RUnlock()
@@ -86,9 +96,11 @@ func (i *dbIter) Last() bool {
 }
 
 func (i *dbIter) Seek(key []byte) bool {
-	if i.p == nil {
+	if i.Released() {
+		i.err = ErrIterReleased
 		return false
 	}
+
 	i.forward = true
 	i.p.mu.RLock()
 	defer i.p.mu.RUnlock()
@@ -100,9 +112,11 @@ func (i *dbIter) Seek(key []byte) bool {
 }
 
 func (i *dbIter) Next() bool {
-	if i.p == nil {
+	if i.Released() {
+		i.err = ErrIterReleased
 		return false
 	}
+
 	if i.node == 0 {
 		if !i.forward {
 			return i.First()
@@ -117,9 +131,11 @@ func (i *dbIter) Next() bool {
 }
 
 func (i *dbIter) Prev() bool {
-	if i.p == nil {
+	if i.Released() {
+		i.err = ErrIterReleased
 		return false
 	}
+
 	if i.node == 0 {
 		if i.forward {
 			return i.Last()
@@ -141,10 +157,10 @@ func (i *dbIter) Value() []byte {
 	return i.value
 }
 
-func (i *dbIter) Error() error { return nil }
+func (i *dbIter) Error() error { return i.err }
 
 func (i *dbIter) Release() {
-	if i.p != nil {
+	if !i.Released() {
 		i.p = nil
 		i.node = 0
 		i.key = nil
@@ -437,6 +453,8 @@ func (p *DB) Reset() {
 // New creates a new initalized in-memory key/value DB. The capacity
 // is the initial key/value buffer capacity. The capacity is advisory,
 // not enforced.
+//
+// The returned DB instance is goroutine-safe.
 func New(cmp comparer.BasicComparer, capacity int) *DB {
 	p := &DB{
 		cmp:       cmp,
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_suite_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_suite_test.go
index 788539a87..18c304b7f 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_suite_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_suite_test.go
@@ -3,15 +3,9 @@ package memdb
 import (
 	"testing"
 
-	. "github.com/onsi/ginkgo"
-	. "github.com/onsi/gomega"
-
 	"github.com/syndtr/goleveldb/leveldb/testutil"
 )
 
-func TestMemdb(t *testing.T) {
-	testutil.RunDefer()
-
-	RegisterFailHandler(Fail)
-	RunSpecs(t, "Memdb Suite")
+func TestMemDB(t *testing.T) {
+	testutil.RunSuite(t, "MemDB Suite")
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_test.go
index f96a9d1ea..5dd6dbc7b 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/memdb/memdb_test.go
@@ -129,7 +129,7 @@ var _ = testutil.Defer(func() {
 				}
 
 				return db
-			})
+			}, nil, nil)
 		})
 	})
 })
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/opt/options.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/opt/options.go
index b940ce427..61f0eadf9 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/opt/options.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/opt/options.go
@@ -11,6 +11,7 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/cache"
 	"github.com/syndtr/goleveldb/leveldb/comparer"
 	"github.com/syndtr/goleveldb/leveldb/filter"
+	"math"
 )
 
 const (
@@ -19,25 +20,57 @@ const (
 	GiB = MiB * 1024
 )
 
-const (
-	DefaultBlockCacheSize       = 8 * MiB
-	DefaultBlockRestartInterval = 16
-	DefaultBlockSize            = 4 * KiB
-	DefaultCompressionType      = SnappyCompression
-	DefaultMaxOpenFiles         = 1000
-	DefaultWriteBuffer          = 4 * MiB
+var (
+	DefaultBlockCacher                   = LRUCacher
+	DefaultBlockCacheCapacity            = 8 * MiB
+	DefaultBlockRestartInterval          = 16
+	DefaultBlockSize                     = 4 * KiB
+	DefaultCompactionExpandLimitFactor   = 25
+	DefaultCompactionGPOverlapsFactor    = 10
+	DefaultCompactionL0Trigger           = 4
+	DefaultCompactionSourceLimitFactor   = 1
+	DefaultCompactionTableSize           = 2 * MiB
+	DefaultCompactionTableSizeMultiplier = 1.0
+	DefaultCompactionTotalSize           = 10 * MiB
+	DefaultCompactionTotalSizeMultiplier = 10.0
+	DefaultCompressionType               = SnappyCompression
+	DefaultIteratorSamplingRate          = 1 * MiB
+	DefaultMaxMemCompationLevel          = 2
+	DefaultNumLevel                      = 7
+	DefaultOpenFilesCacher               = LRUCacher
+	DefaultOpenFilesCacheCapacity        = 500
+	DefaultWriteBuffer                   = 4 * MiB
+	DefaultWriteL0PauseTrigger           = 12
+	DefaultWriteL0SlowdownTrigger        = 8
 )
 
-type noCache struct{}
+// Cacher is a caching algorithm.
+type Cacher interface {
+	New(capacity int) cache.Cacher
+}
+
+type CacherFunc struct {
+	NewFunc func(capacity int) cache.Cacher
+}
+
+func (f *CacherFunc) New(capacity int) cache.Cacher {
+	if f.NewFunc != nil {
+		return f.NewFunc(capacity)
+	}
+	return nil
+}
 
-func (noCache) SetCapacity(capacity int)               {}
-func (noCache) GetNamespace(id uint64) cache.Namespace { return nil }
-func (noCache) Purge(fin cache.PurgeFin)               {}
-func (noCache) Zap(closed bool)                        {}
+func noCacher(int) cache.Cacher { return nil }
 
-var NoCache cache.Cache = noCache{}
+var (
+	// LRUCacher is the LRU-cache algorithm.
+	LRUCacher = &CacherFunc{cache.NewLRU}
 
-// Compression is the per-block compression algorithm to use.
+	// NoCacher is the value to disable caching algorithm.
+	NoCacher = &CacherFunc{}
+)
+
+// Compression is the 'sorted table' block compression algorithm to use.
 type Compression uint
 
 func (c Compression) String() string {
@@ -59,34 +92,47 @@ const (
 	nCompression
 )
 
-// Strict is the DB strict level.
+// Strict is the DB 'strict level'.
 type Strict uint
 
 const (
 	// If present then a corrupted or invalid chunk or block in manifest
-	// journal will cause an error istead of being dropped.
+	// journal will cause an error instead of being dropped.
+	// This will prevent database with corrupted manifest to be opened.
 	StrictManifest Strict = 1 << iota
 
-	// If present then a corrupted or invalid chunk or block in journal
-	// will cause an error istead of being dropped.
-	StrictJournal
-
 	// If present then journal chunk checksum will be verified.
 	StrictJournalChecksum
 
-	// If present then an invalid key/value pair will cause an error
-	// instead of being skipped.
-	StrictIterator
+	// If present then a corrupted or invalid chunk or block in journal
+	// will cause an error instead of being dropped.
+	// This will prevent database with corrupted journal to be opened.
+	StrictJournal
 
 	// If present then 'sorted table' block checksum will be verified.
+	// This has effect on both 'read operation' and compaction.
 	StrictBlockChecksum
 
+	// If present then a corrupted 'sorted table' will fails compaction.
+	// The database will enter read-only mode.
+	StrictCompaction
+
+	// If present then a corrupted 'sorted table' will halts 'read operation'.
+	StrictReader
+
+	// If present then leveldb.Recover will drop corrupted 'sorted table'.
+	StrictRecovery
+
+	// This only applicable for ReadOptions, if present then this ReadOptions
+	// 'strict level' will override global ones.
+	StrictOverride
+
 	// StrictAll enables all strict flags.
-	StrictAll = StrictManifest | StrictJournal | StrictJournalChecksum | StrictIterator | StrictBlockChecksum
+	StrictAll = StrictManifest | StrictJournalChecksum | StrictJournal | StrictBlockChecksum | StrictCompaction | StrictReader | StrictRecovery
 
 	// DefaultStrict is the default strict flags. Specify any strict flags
 	// will override default strict flags as whole (i.e. not OR'ed).
-	DefaultStrict = StrictJournalChecksum | StrictBlockChecksum
+	DefaultStrict = StrictJournalChecksum | StrictBlockChecksum | StrictCompaction | StrictReader
 
 	// NoStrict disables all strict flags. Override default strict flags.
 	NoStrict = ^StrictAll
@@ -101,11 +147,17 @@ type Options struct {
 	// The default value is nil
 	AltFilters []filter.Filter
 
-	// BlockCache provides per-block caching for LevelDB. Specify NoCache to
-	// disable block caching.
+	// BlockCacher provides cache algorithm for LevelDB 'sorted table' block caching.
+	// Specify NoCacher to disable caching algorithm.
 	//
-	// By default LevelDB will create LRU-cache with capacity of 8MiB.
-	BlockCache cache.Cache
+	// The default value is LRUCacher.
+	BlockCacher Cacher
+
+	// BlockCacheCapacity defines the capacity of the 'sorted table' block caching.
+	// Use -1 for zero, this has same effect as specifying NoCacher to BlockCacher.
+	//
+	// The default value is 8MiB.
+	BlockCacheCapacity int
 
 	// BlockRestartInterval is the number of keys between restart points for
 	// delta encoding of keys.
@@ -119,6 +171,73 @@ type Options struct {
 	// The default value is 4KiB.
 	BlockSize int
 
+	// CompactionExpandLimitFactor limits compaction size after expanded.
+	// This will be multiplied by table size limit at compaction target level.
+	//
+	// The default value is 25.
+	CompactionExpandLimitFactor int
+
+	// CompactionGPOverlapsFactor limits overlaps in grandparent (Level + 2) that a
+	// single 'sorted table' generates.
+	// This will be multiplied by table size limit at grandparent level.
+	//
+	// The default value is 10.
+	CompactionGPOverlapsFactor int
+
+	// CompactionL0Trigger defines number of 'sorted table' at level-0 that will
+	// trigger compaction.
+	//
+	// The default value is 4.
+	CompactionL0Trigger int
+
+	// CompactionSourceLimitFactor limits compaction source size. This doesn't apply to
+	// level-0.
+	// This will be multiplied by table size limit at compaction target level.
+	//
+	// The default value is 1.
+	CompactionSourceLimitFactor int
+
+	// CompactionTableSize limits size of 'sorted table' that compaction generates.
+	// The limits for each level will be calculated as:
+	//   CompactionTableSize * (CompactionTableSizeMultiplier ^ Level)
+	// The multiplier for each level can also fine-tuned using CompactionTableSizeMultiplierPerLevel.
+	//
+	// The default value is 2MiB.
+	CompactionTableSize int
+
+	// CompactionTableSizeMultiplier defines multiplier for CompactionTableSize.
+	//
+	// The default value is 1.
+	CompactionTableSizeMultiplier float64
+
+	// CompactionTableSizeMultiplierPerLevel defines per-level multiplier for
+	// CompactionTableSize.
+	// Use zero to skip a level.
+	//
+	// The default value is nil.
+	CompactionTableSizeMultiplierPerLevel []float64
+
+	// CompactionTotalSize limits total size of 'sorted table' for each level.
+	// The limits for each level will be calculated as:
+	//   CompactionTotalSize * (CompactionTotalSizeMultiplier ^ Level)
+	// The multiplier for each level can also fine-tuned using
+	// CompactionTotalSizeMultiplierPerLevel.
+	//
+	// The default value is 10MiB.
+	CompactionTotalSize int
+
+	// CompactionTotalSizeMultiplier defines multiplier for CompactionTotalSize.
+	//
+	// The default value is 10.
+	CompactionTotalSizeMultiplier float64
+
+	// CompactionTotalSizeMultiplierPerLevel defines per-level multiplier for
+	// CompactionTotalSize.
+	// Use zero to skip a level.
+	//
+	// The default value is nil.
+	CompactionTotalSizeMultiplierPerLevel []float64
+
 	// Comparer defines a total ordering over the space of []byte keys: a 'less
 	// than' relationship. The same comparison algorithm must be used for reads
 	// and writes over the lifetime of the DB.
@@ -126,11 +245,22 @@ type Options struct {
 	// The default value uses the same ordering as bytes.Compare.
 	Comparer comparer.Comparer
 
-	// Compression defines the per-block compression to use.
+	// Compression defines the 'sorted table' block compression to use.
 	//
 	// The default value (DefaultCompression) uses snappy compression.
 	Compression Compression
 
+	// DisableBlockCache allows disable use of cache.Cache functionality on
+	// 'sorted table' block.
+	//
+	// The default value is false.
+	DisableBlockCache bool
+
+	// DisableCompactionBackoff allows disable compaction retry backoff.
+	//
+	// The default value is false.
+	DisableCompactionBackoff bool
+
 	// ErrorIfExist defines whether an error should returned if the DB already
 	// exist.
 	//
@@ -159,12 +289,37 @@ type Options struct {
 	// The default value is nil.
 	Filter filter.Filter
 
-	// MaxOpenFiles defines maximum number of open files to kept around
-	// (cached). This is not an hard limit, actual open files may exceed
-	// the defined value.
+	// IteratorSamplingRate defines approximate gap (in bytes) between read
+	// sampling of an iterator. The samples will be used to determine when
+	// compaction should be triggered.
+	//
+	// The default is 1MiB.
+	IteratorSamplingRate int
+
+	// MaxMemCompationLevel defines maximum level a newly compacted 'memdb'
+	// will be pushed into if doesn't creates overlap. This should less than
+	// NumLevel. Use -1 for level-0.
 	//
-	// The default value is 1000.
-	MaxOpenFiles int
+	// The default is 2.
+	MaxMemCompationLevel int
+
+	// NumLevel defines number of database level. The level shouldn't changed
+	// between opens, or the database will panic.
+	//
+	// The default is 7.
+	NumLevel int
+
+	// OpenFilesCacher provides cache algorithm for open files caching.
+	// Specify NoCacher to disable caching algorithm.
+	//
+	// The default value is LRUCacher.
+	OpenFilesCacher Cacher
+
+	// OpenFilesCacheCapacity defines the capacity of the open files caching.
+	// Use -1 for zero, this has same effect as specifying NoCacher to OpenFilesCacher.
+	//
+	// The default value is 500.
+	OpenFilesCacheCapacity int
 
 	// Strict defines the DB strict level.
 	Strict Strict
@@ -177,6 +332,18 @@ type Options struct {
 	//
 	// The default value is 4MiB.
 	WriteBuffer int
+
+	// WriteL0StopTrigger defines number of 'sorted table' at level-0 that will
+	// pause write.
+	//
+	// The default value is 12.
+	WriteL0PauseTrigger int
+
+	// WriteL0SlowdownTrigger defines number of 'sorted table' at level-0 that
+	// will trigger write slowdown.
+	//
+	// The default value is 8.
+	WriteL0SlowdownTrigger int
 }
 
 func (o *Options) GetAltFilters() []filter.Filter {
@@ -186,11 +353,22 @@ func (o *Options) GetAltFilters() []filter.Filter {
 	return o.AltFilters
 }
 
-func (o *Options) GetBlockCache() cache.Cache {
-	if o == nil {
+func (o *Options) GetBlockCacher() Cacher {
+	if o == nil || o.BlockCacher == nil {
+		return DefaultBlockCacher
+	} else if o.BlockCacher == NoCacher {
 		return nil
 	}
-	return o.BlockCache
+	return o.BlockCacher
+}
+
+func (o *Options) GetBlockCacheCapacity() int {
+	if o == nil || o.BlockCacheCapacity == 0 {
+		return DefaultBlockCacheCapacity
+	} else if o.BlockCacheCapacity < 0 {
+		return 0
+	}
+	return o.BlockCacheCapacity
 }
 
 func (o *Options) GetBlockRestartInterval() int {
@@ -207,6 +385,79 @@ func (o *Options) GetBlockSize() int {
 	return o.BlockSize
 }
 
+func (o *Options) GetCompactionExpandLimit(level int) int {
+	factor := DefaultCompactionExpandLimitFactor
+	if o != nil && o.CompactionExpandLimitFactor > 0 {
+		factor = o.CompactionExpandLimitFactor
+	}
+	return o.GetCompactionTableSize(level+1) * factor
+}
+
+func (o *Options) GetCompactionGPOverlaps(level int) int {
+	factor := DefaultCompactionGPOverlapsFactor
+	if o != nil && o.CompactionGPOverlapsFactor > 0 {
+		factor = o.CompactionGPOverlapsFactor
+	}
+	return o.GetCompactionTableSize(level+2) * factor
+}
+
+func (o *Options) GetCompactionL0Trigger() int {
+	if o == nil || o.CompactionL0Trigger == 0 {
+		return DefaultCompactionL0Trigger
+	}
+	return o.CompactionL0Trigger
+}
+
+func (o *Options) GetCompactionSourceLimit(level int) int {
+	factor := DefaultCompactionSourceLimitFactor
+	if o != nil && o.CompactionSourceLimitFactor > 0 {
+		factor = o.CompactionSourceLimitFactor
+	}
+	return o.GetCompactionTableSize(level+1) * factor
+}
+
+func (o *Options) GetCompactionTableSize(level int) int {
+	var (
+		base = DefaultCompactionTableSize
+		mult float64
+	)
+	if o != nil {
+		if o.CompactionTableSize > 0 {
+			base = o.CompactionTableSize
+		}
+		if len(o.CompactionTableSizeMultiplierPerLevel) > level && o.CompactionTableSizeMultiplierPerLevel[level] > 0 {
+			mult = o.CompactionTableSizeMultiplierPerLevel[level]
+		} else if o.CompactionTableSizeMultiplier > 0 {
+			mult = math.Pow(o.CompactionTableSizeMultiplier, float64(level))
+		}
+	}
+	if mult == 0 {
+		mult = math.Pow(DefaultCompactionTableSizeMultiplier, float64(level))
+	}
+	return int(float64(base) * mult)
+}
+
+func (o *Options) GetCompactionTotalSize(level int) int64 {
+	var (
+		base = DefaultCompactionTotalSize
+		mult float64
+	)
+	if o != nil {
+		if o.CompactionTotalSize > 0 {
+			base = o.CompactionTotalSize
+		}
+		if len(o.CompactionTotalSizeMultiplierPerLevel) > level && o.CompactionTotalSizeMultiplierPerLevel[level] > 0 {
+			mult = o.CompactionTotalSizeMultiplierPerLevel[level]
+		} else if o.CompactionTotalSizeMultiplier > 0 {
+			mult = math.Pow(o.CompactionTotalSizeMultiplier, float64(level))
+		}
+	}
+	if mult == 0 {
+		mult = math.Pow(DefaultCompactionTotalSizeMultiplier, float64(level))
+	}
+	return int64(float64(base) * mult)
+}
+
 func (o *Options) GetComparer() comparer.Comparer {
 	if o == nil || o.Comparer == nil {
 		return comparer.DefaultComparer
@@ -221,6 +472,13 @@ func (o *Options) GetCompression() Compression {
 	return o.Compression
 }
 
+func (o *Options) GetDisableCompactionBackoff() bool {
+	if o == nil {
+		return false
+	}
+	return o.DisableCompactionBackoff
+}
+
 func (o *Options) GetErrorIfExist() bool {
 	if o == nil {
 		return false
@@ -242,11 +500,52 @@ func (o *Options) GetFilter() filter.Filter {
 	return o.Filter
 }
 
-func (o *Options) GetMaxOpenFiles() int {
-	if o == nil || o.MaxOpenFiles <= 0 {
-		return DefaultMaxOpenFiles
+func (o *Options) GetIteratorSamplingRate() int {
+	if o == nil || o.IteratorSamplingRate <= 0 {
+		return DefaultIteratorSamplingRate
+	}
+	return o.IteratorSamplingRate
+}
+
+func (o *Options) GetMaxMemCompationLevel() int {
+	level := DefaultMaxMemCompationLevel
+	if o != nil {
+		if o.MaxMemCompationLevel > 0 {
+			level = o.MaxMemCompationLevel
+		} else if o.MaxMemCompationLevel < 0 {
+			level = 0
+		}
+	}
+	if level >= o.GetNumLevel() {
+		return o.GetNumLevel() - 1
+	}
+	return level
+}
+
+func (o *Options) GetNumLevel() int {
+	if o == nil || o.NumLevel <= 0 {
+		return DefaultNumLevel
+	}
+	return o.NumLevel
+}
+
+func (o *Options) GetOpenFilesCacher() Cacher {
+	if o == nil || o.OpenFilesCacher == nil {
+		return DefaultOpenFilesCacher
 	}
-	return o.MaxOpenFiles
+	if o.OpenFilesCacher == NoCacher {
+		return nil
+	}
+	return o.OpenFilesCacher
+}
+
+func (o *Options) GetOpenFilesCacheCapacity() int {
+	if o == nil || o.OpenFilesCacheCapacity == 0 {
+		return DefaultOpenFilesCacheCapacity
+	} else if o.OpenFilesCacheCapacity < 0 {
+		return 0
+	}
+	return o.OpenFilesCacheCapacity
 }
 
 func (o *Options) GetStrict(strict Strict) bool {
@@ -263,6 +562,20 @@ func (o *Options) GetWriteBuffer() int {
 	return o.WriteBuffer
 }
 
+func (o *Options) GetWriteL0PauseTrigger() int {
+	if o == nil || o.WriteL0PauseTrigger == 0 {
+		return DefaultWriteL0PauseTrigger
+	}
+	return o.WriteL0PauseTrigger
+}
+
+func (o *Options) GetWriteL0SlowdownTrigger() int {
+	if o == nil || o.WriteL0SlowdownTrigger == 0 {
+		return DefaultWriteL0SlowdownTrigger
+	}
+	return o.WriteL0SlowdownTrigger
+}
+
 // ReadOptions holds the optional parameters for 'read operation'. The
 // 'read operation' includes Get, Find and NewIterator.
 type ReadOptions struct {
@@ -273,8 +586,8 @@ type ReadOptions struct {
 	// The default value is false.
 	DontFillCache bool
 
-	// Strict overrides global DB strict level. Only StrictIterator and
-	// StrictBlockChecksum that does have effects here.
+	// Strict will be OR'ed with global DB 'strict level' unless StrictOverride
+	// is present. Currently only StrictReader that has effect here.
 	Strict Strict
 }
 
@@ -316,3 +629,11 @@ func (wo *WriteOptions) GetSync() bool {
 	}
 	return wo.Sync
 }
+
+func GetStrict(o *Options, ro *ReadOptions, strict Strict) bool {
+	if ro.GetStrict(StrictOverride) {
+		return ro.GetStrict(strict)
+	} else {
+		return o.GetStrict(strict) || ro.GetStrict(strict)
+	}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/options.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/options.go
index fc6a96965..a3d84ef60 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/options.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/options.go
@@ -7,35 +7,86 @@
 package leveldb
 
 import (
-	"github.com/syndtr/goleveldb/leveldb/cache"
 	"github.com/syndtr/goleveldb/leveldb/filter"
 	"github.com/syndtr/goleveldb/leveldb/opt"
 )
 
-func (s *session) setOptions(o *opt.Options) {
-	s.o = &opt.Options{}
+func dupOptions(o *opt.Options) *opt.Options {
+	newo := &opt.Options{}
 	if o != nil {
-		*s.o = *o
+		*newo = *o
+	}
+	if newo.Strict == 0 {
+		newo.Strict = opt.DefaultStrict
 	}
+	return newo
+}
+
+func (s *session) setOptions(o *opt.Options) {
+	no := dupOptions(o)
 	// Alternative filters.
 	if filters := o.GetAltFilters(); len(filters) > 0 {
-		s.o.AltFilters = make([]filter.Filter, len(filters))
+		no.AltFilters = make([]filter.Filter, len(filters))
 		for i, filter := range filters {
-			s.o.AltFilters[i] = &iFilter{filter}
+			no.AltFilters[i] = &iFilter{filter}
 		}
 	}
-	// Block cache.
-	switch o.GetBlockCache() {
-	case nil:
-		s.o.BlockCache = cache.NewLRUCache(opt.DefaultBlockCacheSize)
-	case opt.NoCache:
-		s.o.BlockCache = nil
-	}
 	// Comparer.
 	s.icmp = &iComparer{o.GetComparer()}
-	s.o.Comparer = s.icmp
+	no.Comparer = s.icmp
 	// Filter.
 	if filter := o.GetFilter(); filter != nil {
-		s.o.Filter = &iFilter{filter}
+		no.Filter = &iFilter{filter}
 	}
+
+	s.o = &cachedOptions{Options: no}
+	s.o.cache()
+}
+
+type cachedOptions struct {
+	*opt.Options
+
+	compactionExpandLimit []int
+	compactionGPOverlaps  []int
+	compactionSourceLimit []int
+	compactionTableSize   []int
+	compactionTotalSize   []int64
+}
+
+func (co *cachedOptions) cache() {
+	numLevel := co.Options.GetNumLevel()
+
+	co.compactionExpandLimit = make([]int, numLevel)
+	co.compactionGPOverlaps = make([]int, numLevel)
+	co.compactionSourceLimit = make([]int, numLevel)
+	co.compactionTableSize = make([]int, numLevel)
+	co.compactionTotalSize = make([]int64, numLevel)
+
+	for level := 0; level < numLevel; level++ {
+		co.compactionExpandLimit[level] = co.Options.GetCompactionExpandLimit(level)
+		co.compactionGPOverlaps[level] = co.Options.GetCompactionGPOverlaps(level)
+		co.compactionSourceLimit[level] = co.Options.GetCompactionSourceLimit(level)
+		co.compactionTableSize[level] = co.Options.GetCompactionTableSize(level)
+		co.compactionTotalSize[level] = co.Options.GetCompactionTotalSize(level)
+	}
+}
+
+func (co *cachedOptions) GetCompactionExpandLimit(level int) int {
+	return co.compactionExpandLimit[level]
+}
+
+func (co *cachedOptions) GetCompactionGPOverlaps(level int) int {
+	return co.compactionGPOverlaps[level]
+}
+
+func (co *cachedOptions) GetCompactionSourceLimit(level int) int {
+	return co.compactionSourceLimit[level]
+}
+
+func (co *cachedOptions) GetCompactionTableSize(level int) int {
+	return co.compactionTableSize[level]
+}
+
+func (co *cachedOptions) GetCompactionTotalSize(level int) int64 {
+	return co.compactionTotalSize[level]
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session.go
index 6b2a61683..b3906f7fc 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session.go
@@ -7,12 +7,13 @@
 package leveldb
 
 import (
-	"errors"
+	"fmt"
 	"io"
 	"os"
 	"sync"
 	"sync/atomic"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/journal"
 	"github.com/syndtr/goleveldb/leveldb/opt"
@@ -20,18 +21,31 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
+type ErrManifestCorrupted struct {
+	Field  string
+	Reason string
+}
+
+func (e *ErrManifestCorrupted) Error() string {
+	return fmt.Sprintf("leveldb: manifest corrupted (field '%s'): %s", e.Field, e.Reason)
+}
+
+func newErrManifestCorrupted(f storage.File, field, reason string) error {
+	return errors.NewErrCorrupted(f, &ErrManifestCorrupted{field, reason})
+}
+
 // session represent a persistent database session.
 type session struct {
 	// Need 64-bit alignment.
-	stFileNum        uint64 // current unused file number
+	stNextFileNum    uint64 // current unused file number
 	stJournalNum     uint64 // current journal file number; need external synchronization
 	stPrevJournalNum uint64 // prev journal file number; no longer used; for compatibility with older version of leveldb
-	stSeq            uint64 // last mem compacted seq; need external synchronization
+	stSeqNum         uint64 // last mem compacted seq; need external synchronization
 	stTempFileNum    uint64
 
 	stor     storage.Storage
 	storLock util.Releaser
-	o        *opt.Options
+	o        *cachedOptions
 	icmp     *iComparer
 	tops     *tOps
 
@@ -39,11 +53,12 @@ type session struct {
 	manifestWriter storage.Writer
 	manifestFile   storage.File
 
-	stCPtrs   [kNumLevels]iKey // compact pointers; need external synchronization
-	stVersion *version         // current version
-	vmu       sync.Mutex
+	stCompPtrs []iKey   // compaction pointers; need external synchronization
+	stVersion  *version // current version
+	vmu        sync.Mutex
 }
 
+// Creates new initialized session instance.
 func newSession(stor storage.Storage, o *opt.Options) (s *session, err error) {
 	if stor == nil {
 		return nil, os.ErrInvalid
@@ -53,22 +68,20 @@ func newSession(stor storage.Storage, o *opt.Options) (s *session, err error) {
 		return
 	}
 	s = &session{
-		stor:     stor,
-		storLock: storLock,
+		stor:       stor,
+		storLock:   storLock,
+		stCompPtrs: make([]iKey, o.GetNumLevel()),
 	}
 	s.setOptions(o)
-	s.tops = newTableOps(s, s.o.GetMaxOpenFiles())
-	s.setVersion(&version{s: s})
-	s.log("log@legend F·NumFile S·FileSize N·Entry C·BadEntry B·BadBlock D·DeletedEntry L·Level Q·SeqNum T·TimeElapsed")
+	s.tops = newTableOps(s)
+	s.setVersion(newVersion(s))
+	s.log("log@legend F·NumFile S·FileSize N·Entry C·BadEntry B·BadBlock Ke·KeyError D·DroppedEntry L·Level Q·SeqNum T·TimeElapsed")
 	return
 }
 
 // Close session.
 func (s *session) close() {
 	s.tops.close()
-	if bc := s.o.GetBlockCache(); bc != nil {
-		bc.Purge(nil)
-	}
 	if s.manifest != nil {
 		s.manifest.Close()
 	}
@@ -81,6 +94,7 @@ func (s *session) close() {
 	s.stVersion = nil
 }
 
+// Release session lock.
 func (s *session) release() {
 	s.storLock.Release()
 }
@@ -98,26 +112,26 @@ func (s *session) recover() (err error) {
 			// Don't return os.ErrNotExist if the underlying storage contains
 			// other files that belong to LevelDB. So the DB won't get trashed.
 			if files, _ := s.stor.GetFiles(storage.TypeAll); len(files) > 0 {
-				err = ErrCorrupted{Type: CorruptedManifest, Err: errors.New("leveldb: manifest file missing")}
+				err = &errors.ErrCorrupted{File: &storage.FileInfo{Type: storage.TypeManifest}, Err: &errors.ErrMissingFiles{}}
 			}
 		}
 	}()
 
-	file, err := s.stor.GetManifest()
+	m, err := s.stor.GetManifest()
 	if err != nil {
 		return
 	}
 
-	reader, err := file.Open()
+	reader, err := m.Open()
 	if err != nil {
 		return
 	}
 	defer reader.Close()
 	strict := s.o.GetStrict(opt.StrictManifest)
-	jr := journal.NewReader(reader, dropper{s, file}, strict, true)
+	jr := journal.NewReader(reader, dropper{s, m}, strict, true)
 
-	staging := s.version_NB().newStaging()
-	rec := &sessionRecord{}
+	staging := s.stVersion.newStaging()
+	rec := &sessionRecord{numLevel: s.o.GetNumLevel()}
 	for {
 		var r io.Reader
 		r, err = jr.Next()
@@ -126,51 +140,57 @@ func (s *session) recover() (err error) {
 				err = nil
 				break
 			}
-			return
+			return errors.SetFile(err, m)
 		}
 
 		err = rec.decode(r)
 		if err == nil {
 			// save compact pointers
-			for _, rp := range rec.compactionPointers {
-				s.stCPtrs[rp.level] = iKey(rp.key)
+			for _, r := range rec.compPtrs {
+				s.stCompPtrs[r.level] = iKey(r.ikey)
 			}
 			// commit record to version staging
 			staging.commit(rec)
-		} else if strict {
-			return ErrCorrupted{Type: CorruptedManifest, Err: err}
 		} else {
-			s.logf("manifest error: %v (skipped)", err)
+			err = errors.SetFile(err, m)
+			if strict || !errors.IsCorrupted(err) {
+				return
+			} else {
+				s.logf("manifest error: %v (skipped)", errors.SetFile(err, m))
+			}
 		}
-		rec.resetCompactionPointers()
+		rec.resetCompPtrs()
 		rec.resetAddedTables()
 		rec.resetDeletedTables()
 	}
 
 	switch {
 	case !rec.has(recComparer):
-		return ErrCorrupted{Type: CorruptedManifest, Err: errors.New("leveldb: manifest missing comparer name")}
+		return newErrManifestCorrupted(m, "comparer", "missing")
 	case rec.comparer != s.icmp.uName():
-		return ErrCorrupted{Type: CorruptedManifest, Err: errors.New("leveldb: comparer mismatch, " + "want '" + s.icmp.uName() + "', " + "got '" + rec.comparer + "'")}
-	case !rec.has(recNextNum):
-		return ErrCorrupted{Type: CorruptedManifest, Err: errors.New("leveldb: manifest missing next file number")}
+		return newErrManifestCorrupted(m, "comparer", fmt.Sprintf("mismatch: want '%s', got '%s'", s.icmp.uName(), rec.comparer))
+	case !rec.has(recNextFileNum):
+		return newErrManifestCorrupted(m, "next-file-num", "missing")
 	case !rec.has(recJournalNum):
-		return ErrCorrupted{Type: CorruptedManifest, Err: errors.New("leveldb: manifest missing journal file number")}
-	case !rec.has(recSeq):
-		return ErrCorrupted{Type: CorruptedManifest, Err: errors.New("leveldb: manifest missing seq number")}
+		return newErrManifestCorrupted(m, "journal-file-num", "missing")
+	case !rec.has(recSeqNum):
+		return newErrManifestCorrupted(m, "seq-num", "missing")
 	}
 
-	s.manifestFile = file
+	s.manifestFile = m
 	s.setVersion(staging.finish())
-	s.setFileNum(rec.nextNum)
+	s.setNextFileNum(rec.nextFileNum)
 	s.recordCommited(rec)
 	return nil
 }
 
 // Commit session; need external synchronization.
 func (s *session) commit(r *sessionRecord) (err error) {
+	v := s.version()
+	defer v.release()
+
 	// spawn new version based on current version
-	nv := s.version_NB().spawn(r)
+	nv := v.spawn(r)
 
 	if s.manifest == nil {
 		// manifest journal writer not yet created, create one
@@ -189,22 +209,22 @@ func (s *session) commit(r *sessionRecord) (err error) {
 
 // Pick a compaction based on current state; need external synchronization.
 func (s *session) pickCompaction() *compaction {
-	v := s.version_NB()
+	v := s.version()
 
 	var level int
 	var t0 tFiles
 	if v.cScore >= 1 {
 		level = v.cLevel
-		cp := s.stCPtrs[level]
-		tt := v.tables[level]
-		for _, t := range tt {
-			if cp == nil || s.icmp.Compare(t.max, cp) > 0 {
+		cptr := s.stCompPtrs[level]
+		tables := v.tables[level]
+		for _, t := range tables {
+			if cptr == nil || s.icmp.Compare(t.imax, cptr) > 0 {
 				t0 = append(t0, t)
 				break
 			}
 		}
 		if len(t0) == 0 {
-			t0 = append(t0, tt[0])
+			t0 = append(t0, tables[0])
 		}
 	} else {
 		if p := atomic.LoadPointer(&v.cSeek); p != nil {
@@ -212,29 +232,21 @@ func (s *session) pickCompaction() *compaction {
 			level = ts.level
 			t0 = append(t0, ts.table)
 		} else {
+			v.release()
 			return nil
 		}
 	}
 
-	c := &compaction{s: s, version: v, level: level}
-	if level == 0 {
-		min, max := t0.getRange(s.icmp)
-		t0 = nil
-		v.tables[0].getOverlaps(min.ukey(), max.ukey(), &t0, false, s.icmp.ucmp)
-	}
-
-	c.tables[0] = t0
-	c.expand()
-	return c
+	return newCompaction(s, v, level, t0)
 }
 
 // Create compaction from given level and range; need external synchronization.
-func (s *session) getCompactionRange(level int, min, max []byte) *compaction {
-	v := s.version_NB()
+func (s *session) getCompactionRange(level int, umin, umax []byte) *compaction {
+	v := s.version()
 
-	var t0 tFiles
-	v.tables[level].getOverlaps(min, max, &t0, level != 0, s.icmp.ucmp)
+	t0 := v.tables[level].getOverlaps(nil, s.icmp, umin, umax, level == 0)
 	if len(t0) == 0 {
+		v.release()
 		return nil
 	}
 
@@ -243,7 +255,7 @@ func (s *session) getCompactionRange(level int, min, max []byte) *compaction {
 	// and we must not pick one file and drop another older file if the
 	// two files overlap.
 	if level > 0 {
-		limit := uint64(kMaxTableSize)
+		limit := uint64(v.s.o.GetCompactionSourceLimit(level))
 		total := uint64(0)
 		for i, t := range t0 {
 			total += t.size
@@ -255,90 +267,124 @@ func (s *session) getCompactionRange(level int, min, max []byte) *compaction {
 		}
 	}
 
-	c := &compaction{s: s, version: v, level: level}
-	c.tables[0] = t0
+	return newCompaction(s, v, level, t0)
+}
+
+func newCompaction(s *session, v *version, level int, t0 tFiles) *compaction {
+	c := &compaction{
+		s:             s,
+		v:             v,
+		level:         level,
+		tables:        [2]tFiles{t0, nil},
+		maxGPOverlaps: uint64(s.o.GetCompactionGPOverlaps(level)),
+		tPtrs:         make([]int, s.o.GetNumLevel()),
+	}
 	c.expand()
+	c.save()
 	return c
 }
 
-// compaction represent a compaction state
+// compaction represent a compaction state.
 type compaction struct {
-	s       *session
-	version *version
+	s *session
+	v *version
+
+	level         int
+	tables        [2]tFiles
+	maxGPOverlaps uint64
+
+	gp                tFiles
+	gpi               int
+	seenKey           bool
+	gpOverlappedBytes uint64
+	imin, imax        iKey
+	tPtrs             []int
+	released          bool
+
+	snapGPI               int
+	snapSeenKey           bool
+	snapGPOverlappedBytes uint64
+	snapTPtrs             []int
+}
 
-	level  int
-	tables [2]tFiles
+func (c *compaction) save() {
+	c.snapGPI = c.gpi
+	c.snapSeenKey = c.seenKey
+	c.snapGPOverlappedBytes = c.gpOverlappedBytes
+	c.snapTPtrs = append(c.snapTPtrs[:0], c.tPtrs...)
+}
 
-	gp              tFiles
-	gpidx           int
-	seenKey         bool
-	overlappedBytes uint64
-	min, max        iKey
+func (c *compaction) restore() {
+	c.gpi = c.snapGPI
+	c.seenKey = c.snapSeenKey
+	c.gpOverlappedBytes = c.snapGPOverlappedBytes
+	c.tPtrs = append(c.tPtrs[:0], c.snapTPtrs...)
+}
 
-	tPtrs [kNumLevels]int
+func (c *compaction) release() {
+	if !c.released {
+		c.released = true
+		c.v.release()
+	}
 }
 
 // Expand compacted tables; need external synchronization.
 func (c *compaction) expand() {
-	s := c.s
-	v := c.version
-
-	level := c.level
-	vt0, vt1 := v.tables[level], v.tables[level+1]
+	limit := uint64(c.s.o.GetCompactionExpandLimit(c.level))
+	vt0, vt1 := c.v.tables[c.level], c.v.tables[c.level+1]
 
 	t0, t1 := c.tables[0], c.tables[1]
-	min, max := t0.getRange(s.icmp)
-	vt1.getOverlaps(min.ukey(), max.ukey(), &t1, true, s.icmp.ucmp)
-
-	// Get entire range covered by compaction
-	amin, amax := append(t0, t1...).getRange(s.icmp)
+	imin, imax := t0.getRange(c.s.icmp)
+	// We expand t0 here just incase ukey hop across tables.
+	t0 = vt0.getOverlaps(t0, c.s.icmp, imin.ukey(), imax.ukey(), c.level == 0)
+	if len(t0) != len(c.tables[0]) {
+		imin, imax = t0.getRange(c.s.icmp)
+	}
+	t1 = vt1.getOverlaps(t1, c.s.icmp, imin.ukey(), imax.ukey(), false)
+	// Get entire range covered by compaction.
+	amin, amax := append(t0, t1...).getRange(c.s.icmp)
 
 	// See if we can grow the number of inputs in "level" without
 	// changing the number of "level+1" files we pick up.
 	if len(t1) > 0 {
-		var exp0 tFiles
-		vt0.getOverlaps(amin.ukey(), amax.ukey(), &exp0, level != 0, s.icmp.ucmp)
-		if len(exp0) > len(t0) && t1.size()+exp0.size() < kExpCompactionMaxBytes {
-			var exp1 tFiles
-			xmin, xmax := exp0.getRange(s.icmp)
-			vt1.getOverlaps(xmin.ukey(), xmax.ukey(), &exp1, true, s.icmp.ucmp)
+		exp0 := vt0.getOverlaps(nil, c.s.icmp, amin.ukey(), amax.ukey(), c.level == 0)
+		if len(exp0) > len(t0) && t1.size()+exp0.size() < limit {
+			xmin, xmax := exp0.getRange(c.s.icmp)
+			exp1 := vt1.getOverlaps(nil, c.s.icmp, xmin.ukey(), xmax.ukey(), false)
 			if len(exp1) == len(t1) {
-				s.logf("table@compaction expanding L%d+L%d (F·%d S·%s)+(F·%d S·%s) -> (F·%d S·%s)+(F·%d S·%s)",
-					level, level+1, len(t0), shortenb(int(t0.size())), len(t1), shortenb(int(t1.size())),
+				c.s.logf("table@compaction expanding L%d+L%d (F·%d S·%s)+(F·%d S·%s) -> (F·%d S·%s)+(F·%d S·%s)",
+					c.level, c.level+1, len(t0), shortenb(int(t0.size())), len(t1), shortenb(int(t1.size())),
 					len(exp0), shortenb(int(exp0.size())), len(exp1), shortenb(int(exp1.size())))
-				min, max = xmin, xmax
+				imin, imax = xmin, xmax
 				t0, t1 = exp0, exp1
-				amin, amax = append(t0, t1...).getRange(s.icmp)
+				amin, amax = append(t0, t1...).getRange(c.s.icmp)
 			}
 		}
 	}
 
 	// Compute the set of grandparent files that overlap this compaction
 	// (parent == level+1; grandparent == level+2)
-	if level+2 < kNumLevels {
-		v.tables[level+2].getOverlaps(amin.ukey(), amax.ukey(), &c.gp, true, s.icmp.ucmp)
+	if c.level+2 < c.s.o.GetNumLevel() {
+		c.gp = c.v.tables[c.level+2].getOverlaps(c.gp, c.s.icmp, amin.ukey(), amax.ukey(), false)
 	}
 
 	c.tables[0], c.tables[1] = t0, t1
-	c.min, c.max = min, max
+	c.imin, c.imax = imin, imax
 }
 
 // Check whether compaction is trivial.
 func (c *compaction) trivial() bool {
-	return len(c.tables[0]) == 1 && len(c.tables[1]) == 0 && c.gp.size() <= kMaxGrandParentOverlapBytes
+	return len(c.tables[0]) == 1 && len(c.tables[1]) == 0 && c.gp.size() <= c.maxGPOverlaps
 }
 
-func (c *compaction) isBaseLevelForKey(key []byte) bool {
-	s := c.s
-	v := c.version
-
-	for level, tt := range v.tables[c.level+2:] {
-		for c.tPtrs[level] < len(tt) {
-			t := tt[c.tPtrs[level]]
-			if s.icmp.uCompare(key, t.max.ukey()) <= 0 {
-				// We've advanced far enough
-				if s.icmp.uCompare(key, t.min.ukey()) >= 0 {
-					// Key falls in this file's range, so definitely not base level
+func (c *compaction) baseLevelForKey(ukey []byte) bool {
+	for level, tables := range c.v.tables[c.level+2:] {
+		for c.tPtrs[level] < len(tables) {
+			t := tables[c.tPtrs[level]]
+			if c.s.icmp.uCompare(ukey, t.imax.ukey()) <= 0 {
+				// We've advanced far enough.
+				if c.s.icmp.uCompare(ukey, t.imin.ukey()) >= 0 {
+					// Key falls in this file's range, so definitely not base level.
 					return false
 				}
 				break
@@ -349,55 +395,61 @@ func (c *compaction) isBaseLevelForKey(key []byte) bool {
 	return true
 }
 
-func (c *compaction) shouldStopBefore(key iKey) bool {
-	for ; c.gpidx < len(c.gp); c.gpidx++ {
-		gp := c.gp[c.gpidx]
-		if c.s.icmp.Compare(key, gp.max) <= 0 {
+func (c *compaction) shouldStopBefore(ikey iKey) bool {
+	for ; c.gpi < len(c.gp); c.gpi++ {
+		gp := c.gp[c.gpi]
+		if c.s.icmp.Compare(ikey, gp.imax) <= 0 {
 			break
 		}
 		if c.seenKey {
-			c.overlappedBytes += gp.size
+			c.gpOverlappedBytes += gp.size
 		}
 	}
 	c.seenKey = true
 
-	if c.overlappedBytes > kMaxGrandParentOverlapBytes {
-		// Too much overlap for current output; start new output
-		c.overlappedBytes = 0
+	if c.gpOverlappedBytes > c.maxGPOverlaps {
+		// Too much overlap for current output; start new output.
+		c.gpOverlappedBytes = 0
 		return true
 	}
 	return false
 }
 
+// Creates an iterator.
 func (c *compaction) newIterator() iterator.Iterator {
-	s := c.s
-
-	level := c.level
-	icap := 2
+	// Creates iterator slice.
+	icap := len(c.tables)
 	if c.level == 0 {
+		// Special case for level-0
 		icap = len(c.tables[0]) + 1
 	}
 	its := make([]iterator.Iterator, 0, icap)
 
+	// Options.
 	ro := &opt.ReadOptions{
 		DontFillCache: true,
+		Strict:        opt.StrictOverride,
+	}
+	strict := c.s.o.GetStrict(opt.StrictCompaction)
+	if strict {
+		ro.Strict |= opt.StrictReader
 	}
-	strict := s.o.GetStrict(opt.StrictIterator)
 
-	for i, tt := range c.tables {
-		if len(tt) == 0 {
+	for i, tables := range c.tables {
+		if len(tables) == 0 {
 			continue
 		}
 
-		if level+i == 0 {
-			for _, t := range tt {
-				its = append(its, s.tops.newIterator(t, nil, ro))
+		// Level-0 is not sorted and may overlaps each other.
+		if c.level+i == 0 {
+			for _, t := range tables {
+				its = append(its, c.s.tops.newIterator(t, nil, ro))
 			}
 		} else {
-			it := iterator.NewIndexedIterator(tt.newIndexIterator(s.tops, s.icmp, nil, ro), strict, true)
+			it := iterator.NewIndexedIterator(tables.newIndexIterator(c.s.tops, c.s.icmp, nil, ro), strict)
 			its = append(its, it)
 		}
 	}
 
-	return iterator.NewMergedIterator(its, s.icmp, true)
+	return iterator.NewMergedIterator(its, c.s.icmp, strict)
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record.go
index c50fda737..1bdcc68f5 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record.go
@@ -9,11 +9,11 @@ package leveldb
 import (
 	"bufio"
 	"encoding/binary"
-	"errors"
 	"io"
-)
+	"strings"
 
-var errCorruptManifest = errors.New("leveldb: corrupt manifest")
+	"github.com/syndtr/goleveldb/leveldb/errors"
+)
 
 type byteReader interface {
 	io.Reader
@@ -22,32 +22,28 @@ type byteReader interface {
 
 // These numbers are written to disk and should not be changed.
 const (
-	recComparer          = 1
-	recJournalNum        = 2
-	recNextNum           = 3
-	recSeq               = 4
-	recCompactionPointer = 5
-	recDeletedTable      = 6
-	recNewTable          = 7
+	recComparer    = 1
+	recJournalNum  = 2
+	recNextFileNum = 3
+	recSeqNum      = 4
+	recCompPtr     = 5
+	recDelTable    = 6
+	recAddTable    = 7
 	// 8 was used for large value refs
 	recPrevJournalNum = 9
 )
 
 type cpRecord struct {
 	level int
-	key   iKey
+	ikey  iKey
 }
 
-type ntRecord struct {
+type atRecord struct {
 	level int
 	num   uint64
 	size  uint64
-	min   iKey
-	max   iKey
-}
-
-func (r ntRecord) makeFile(s *session) *tFile {
-	return newTFile(s.getTableFile(r.num), r.size, r.min, r.max)
+	imin  iKey
+	imax  iKey
 }
 
 type dtRecord struct {
@@ -56,17 +52,20 @@ type dtRecord struct {
 }
 
 type sessionRecord struct {
-	hasRec             int
-	comparer           string
-	journalNum         uint64
-	prevJournalNum     uint64
-	nextNum            uint64
-	seq                uint64
-	compactionPointers []cpRecord
-	addedTables        []ntRecord
-	deletedTables      []dtRecord
-	scratch            [binary.MaxVarintLen64]byte
-	err                error
+	numLevel int
+
+	hasRec         int
+	comparer       string
+	journalNum     uint64
+	prevJournalNum uint64
+	nextFileNum    uint64
+	seqNum         uint64
+	compPtrs       []cpRecord
+	addedTables    []atRecord
+	deletedTables  []dtRecord
+
+	scratch [binary.MaxVarintLen64]byte
+	err     error
 }
 
 func (p *sessionRecord) has(rec int) bool {
@@ -88,47 +87,47 @@ func (p *sessionRecord) setPrevJournalNum(num uint64) {
 	p.prevJournalNum = num
 }
 
-func (p *sessionRecord) setNextNum(num uint64) {
-	p.hasRec |= 1 << recNextNum
-	p.nextNum = num
+func (p *sessionRecord) setNextFileNum(num uint64) {
+	p.hasRec |= 1 << recNextFileNum
+	p.nextFileNum = num
 }
 
-func (p *sessionRecord) setSeq(seq uint64) {
-	p.hasRec |= 1 << recSeq
-	p.seq = seq
+func (p *sessionRecord) setSeqNum(num uint64) {
+	p.hasRec |= 1 << recSeqNum
+	p.seqNum = num
 }
 
-func (p *sessionRecord) addCompactionPointer(level int, key iKey) {
-	p.hasRec |= 1 << recCompactionPointer
-	p.compactionPointers = append(p.compactionPointers, cpRecord{level, key})
+func (p *sessionRecord) addCompPtr(level int, ikey iKey) {
+	p.hasRec |= 1 << recCompPtr
+	p.compPtrs = append(p.compPtrs, cpRecord{level, ikey})
 }
 
-func (p *sessionRecord) resetCompactionPointers() {
-	p.hasRec &= ^(1 << recCompactionPointer)
-	p.compactionPointers = p.compactionPointers[:0]
+func (p *sessionRecord) resetCompPtrs() {
+	p.hasRec &= ^(1 << recCompPtr)
+	p.compPtrs = p.compPtrs[:0]
 }
 
-func (p *sessionRecord) addTable(level int, num, size uint64, min, max iKey) {
-	p.hasRec |= 1 << recNewTable
-	p.addedTables = append(p.addedTables, ntRecord{level, num, size, min, max})
+func (p *sessionRecord) addTable(level int, num, size uint64, imin, imax iKey) {
+	p.hasRec |= 1 << recAddTable
+	p.addedTables = append(p.addedTables, atRecord{level, num, size, imin, imax})
 }
 
 func (p *sessionRecord) addTableFile(level int, t *tFile) {
-	p.addTable(level, t.file.Num(), t.size, t.min, t.max)
+	p.addTable(level, t.file.Num(), t.size, t.imin, t.imax)
 }
 
 func (p *sessionRecord) resetAddedTables() {
-	p.hasRec &= ^(1 << recNewTable)
+	p.hasRec &= ^(1 << recAddTable)
 	p.addedTables = p.addedTables[:0]
 }
 
-func (p *sessionRecord) deleteTable(level int, num uint64) {
-	p.hasRec |= 1 << recDeletedTable
+func (p *sessionRecord) delTable(level int, num uint64) {
+	p.hasRec |= 1 << recDelTable
 	p.deletedTables = append(p.deletedTables, dtRecord{level, num})
 }
 
 func (p *sessionRecord) resetDeletedTables() {
-	p.hasRec &= ^(1 << recDeletedTable)
+	p.hasRec &= ^(1 << recDelTable)
 	p.deletedTables = p.deletedTables[:0]
 }
 
@@ -161,43 +160,45 @@ func (p *sessionRecord) encode(w io.Writer) error {
 		p.putUvarint(w, recJournalNum)
 		p.putUvarint(w, p.journalNum)
 	}
-	if p.has(recNextNum) {
-		p.putUvarint(w, recNextNum)
-		p.putUvarint(w, p.nextNum)
+	if p.has(recNextFileNum) {
+		p.putUvarint(w, recNextFileNum)
+		p.putUvarint(w, p.nextFileNum)
 	}
-	if p.has(recSeq) {
-		p.putUvarint(w, recSeq)
-		p.putUvarint(w, p.seq)
+	if p.has(recSeqNum) {
+		p.putUvarint(w, recSeqNum)
+		p.putUvarint(w, p.seqNum)
 	}
-	for _, cp := range p.compactionPointers {
-		p.putUvarint(w, recCompactionPointer)
-		p.putUvarint(w, uint64(cp.level))
-		p.putBytes(w, cp.key)
+	for _, r := range p.compPtrs {
+		p.putUvarint(w, recCompPtr)
+		p.putUvarint(w, uint64(r.level))
+		p.putBytes(w, r.ikey)
 	}
-	for _, t := range p.deletedTables {
-		p.putUvarint(w, recDeletedTable)
-		p.putUvarint(w, uint64(t.level))
-		p.putUvarint(w, t.num)
+	for _, r := range p.deletedTables {
+		p.putUvarint(w, recDelTable)
+		p.putUvarint(w, uint64(r.level))
+		p.putUvarint(w, r.num)
 	}
-	for _, t := range p.addedTables {
-		p.putUvarint(w, recNewTable)
-		p.putUvarint(w, uint64(t.level))
-		p.putUvarint(w, t.num)
-		p.putUvarint(w, t.size)
-		p.putBytes(w, t.min)
-		p.putBytes(w, t.max)
+	for _, r := range p.addedTables {
+		p.putUvarint(w, recAddTable)
+		p.putUvarint(w, uint64(r.level))
+		p.putUvarint(w, r.num)
+		p.putUvarint(w, r.size)
+		p.putBytes(w, r.imin)
+		p.putBytes(w, r.imax)
 	}
 	return p.err
 }
 
-func (p *sessionRecord) readUvarint(r io.ByteReader) uint64 {
+func (p *sessionRecord) readUvarintMayEOF(field string, r io.ByteReader, mayEOF bool) uint64 {
 	if p.err != nil {
 		return 0
 	}
 	x, err := binary.ReadUvarint(r)
 	if err != nil {
-		if err == io.EOF {
-			p.err = errCorruptManifest
+		if err == io.ErrUnexpectedEOF || (mayEOF == false && err == io.EOF) {
+			p.err = errors.NewErrCorrupted(nil, &ErrManifestCorrupted{field, "short read"})
+		} else if strings.HasPrefix(err.Error(), "binary:") {
+			p.err = errors.NewErrCorrupted(nil, &ErrManifestCorrupted{field, err.Error()})
 		} else {
 			p.err = err
 		}
@@ -206,35 +207,39 @@ func (p *sessionRecord) readUvarint(r io.ByteReader) uint64 {
 	return x
 }
 
-func (p *sessionRecord) readBytes(r byteReader) []byte {
+func (p *sessionRecord) readUvarint(field string, r io.ByteReader) uint64 {
+	return p.readUvarintMayEOF(field, r, false)
+}
+
+func (p *sessionRecord) readBytes(field string, r byteReader) []byte {
 	if p.err != nil {
 		return nil
 	}
-	n := p.readUvarint(r)
+	n := p.readUvarint(field, r)
 	if p.err != nil {
 		return nil
 	}
 	x := make([]byte, n)
 	_, p.err = io.ReadFull(r, x)
 	if p.err != nil {
-		if p.err == io.EOF {
-			p.err = errCorruptManifest
+		if p.err == io.ErrUnexpectedEOF {
+			p.err = errors.NewErrCorrupted(nil, &ErrManifestCorrupted{field, "short read"})
 		}
 		return nil
 	}
 	return x
 }
 
-func (p *sessionRecord) readLevel(r io.ByteReader) int {
+func (p *sessionRecord) readLevel(field string, r io.ByteReader) int {
 	if p.err != nil {
 		return 0
 	}
-	x := p.readUvarint(r)
+	x := p.readUvarint(field, r)
 	if p.err != nil {
 		return 0
 	}
-	if x >= kNumLevels {
-		p.err = errCorruptManifest
+	if x >= uint64(p.numLevel) {
+		p.err = errors.NewErrCorrupted(nil, &ErrManifestCorrupted{field, "invalid level number"})
 		return 0
 	}
 	return int(x)
@@ -247,59 +252,59 @@ func (p *sessionRecord) decode(r io.Reader) error {
 	}
 	p.err = nil
 	for p.err == nil {
-		rec, err := binary.ReadUvarint(br)
-		if err != nil {
-			if err == io.EOF {
-				err = nil
+		rec := p.readUvarintMayEOF("field-header", br, true)
+		if p.err != nil {
+			if p.err == io.EOF {
+				return nil
 			}
-			return err
+			return p.err
 		}
 		switch rec {
 		case recComparer:
-			x := p.readBytes(br)
+			x := p.readBytes("comparer", br)
 			if p.err == nil {
 				p.setComparer(string(x))
 			}
 		case recJournalNum:
-			x := p.readUvarint(br)
+			x := p.readUvarint("journal-num", br)
 			if p.err == nil {
 				p.setJournalNum(x)
 			}
 		case recPrevJournalNum:
-			x := p.readUvarint(br)
+			x := p.readUvarint("prev-journal-num", br)
 			if p.err == nil {
 				p.setPrevJournalNum(x)
 			}
-		case recNextNum:
-			x := p.readUvarint(br)
+		case recNextFileNum:
+			x := p.readUvarint("next-file-num", br)
 			if p.err == nil {
-				p.setNextNum(x)
+				p.setNextFileNum(x)
 			}
-		case recSeq:
-			x := p.readUvarint(br)
+		case recSeqNum:
+			x := p.readUvarint("seq-num", br)
 			if p.err == nil {
-				p.setSeq(x)
+				p.setSeqNum(x)
 			}
-		case recCompactionPointer:
-			level := p.readLevel(br)
-			key := p.readBytes(br)
+		case recCompPtr:
+			level := p.readLevel("comp-ptr.level", br)
+			ikey := p.readBytes("comp-ptr.ikey", br)
 			if p.err == nil {
-				p.addCompactionPointer(level, iKey(key))
+				p.addCompPtr(level, iKey(ikey))
 			}
-		case recNewTable:
-			level := p.readLevel(br)
-			num := p.readUvarint(br)
-			size := p.readUvarint(br)
-			min := p.readBytes(br)
-			max := p.readBytes(br)
+		case recAddTable:
+			level := p.readLevel("add-table.level", br)
+			num := p.readUvarint("add-table.num", br)
+			size := p.readUvarint("add-table.size", br)
+			imin := p.readBytes("add-table.imin", br)
+			imax := p.readBytes("add-table.imax", br)
 			if p.err == nil {
-				p.addTable(level, num, size, min, max)
+				p.addTable(level, num, size, imin, imax)
 			}
-		case recDeletedTable:
-			level := p.readLevel(br)
-			num := p.readUvarint(br)
+		case recDelTable:
+			level := p.readLevel("del-table.level", br)
+			num := p.readUvarint("del-table.num", br)
 			if p.err == nil {
-				p.deleteTable(level, num)
+				p.delTable(level, num)
 			}
 		}
 	}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record_test.go
index 029fabfe6..c0c035ae3 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_record_test.go
@@ -9,6 +9,8 @@ package leveldb
 import (
 	"bytes"
 	"testing"
+
+	"github.com/syndtr/goleveldb/leveldb/opt"
 )
 
 func decodeEncode(v *sessionRecord) (res bool, err error) {
@@ -17,7 +19,7 @@ func decodeEncode(v *sessionRecord) (res bool, err error) {
 	if err != nil {
 		return
 	}
-	v2 := new(sessionRecord)
+	v2 := &sessionRecord{numLevel: opt.DefaultNumLevel}
 	err = v.decode(b)
 	if err != nil {
 		return
@@ -32,7 +34,7 @@ func decodeEncode(v *sessionRecord) (res bool, err error) {
 
 func TestSessionRecord_EncodeDecode(t *testing.T) {
 	big := uint64(1) << 50
-	v := new(sessionRecord)
+	v := &sessionRecord{numLevel: opt.DefaultNumLevel}
 	i := uint64(0)
 	test := func() {
 		res, err := decodeEncode(v)
@@ -47,16 +49,16 @@ func TestSessionRecord_EncodeDecode(t *testing.T) {
 	for ; i < 4; i++ {
 		test()
 		v.addTable(3, big+300+i, big+400+i,
-			newIKey([]byte("foo"), big+500+1, tVal),
-			newIKey([]byte("zoo"), big+600+1, tDel))
-		v.deleteTable(4, big+700+i)
-		v.addCompactionPointer(int(i), newIKey([]byte("x"), big+900+1, tVal))
+			newIkey([]byte("foo"), big+500+1, ktVal),
+			newIkey([]byte("zoo"), big+600+1, ktDel))
+		v.delTable(4, big+700+i)
+		v.addCompPtr(int(i), newIkey([]byte("x"), big+900+1, ktVal))
 	}
 
 	v.setComparer("foo")
 	v.setJournalNum(big + 100)
 	v.setPrevJournalNum(big + 99)
-	v.setNextNum(big + 200)
-	v.setSeq(big + 1000)
+	v.setNextFileNum(big + 200)
+	v.setSeqNum(big + 1000)
 	test()
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_util.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_util.go
index bf412b030..007c02cde 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_util.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/session_util.go
@@ -14,7 +14,7 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/storage"
 )
 
-// logging
+// Logging.
 
 type dropper struct {
 	s    *session
@@ -22,22 +22,17 @@ type dropper struct {
 }
 
 func (d dropper) Drop(err error) {
-	if e, ok := err.(journal.DroppedError); ok {
+	if e, ok := err.(*journal.ErrCorrupted); ok {
 		d.s.logf("journal@drop %s-%d S·%s %q", d.file.Type(), d.file.Num(), shortenb(e.Size), e.Reason)
 	} else {
 		d.s.logf("journal@drop %s-%d %q", d.file.Type(), d.file.Num(), err)
 	}
 }
 
-func (s *session) log(v ...interface{}) {
-	s.stor.Log(fmt.Sprint(v...))
-}
-
-func (s *session) logf(format string, v ...interface{}) {
-	s.stor.Log(fmt.Sprintf(format, v...))
-}
+func (s *session) log(v ...interface{})                 { s.stor.Log(fmt.Sprint(v...)) }
+func (s *session) logf(format string, v ...interface{}) { s.stor.Log(fmt.Sprintf(format, v...)) }
 
-// file utils
+// File utils.
 
 func (s *session) getJournalFile(num uint64) storage.File {
 	return s.stor.GetFile(num, storage.TypeJournal)
@@ -56,9 +51,14 @@ func (s *session) newTemp() storage.File {
 	return s.stor.GetFile(num, storage.TypeTemp)
 }
 
-// session state
+func (s *session) tableFileFromRecord(r atRecord) *tFile {
+	return newTableFile(s.getTableFile(r.num), r.size, r.imin, r.imax)
+}
+
+// Session state.
 
-// Get current version.
+// Get current version. This will incr version ref, must call
+// version.release (exactly once) after use.
 func (s *session) version() *version {
 	s.vmu.Lock()
 	defer s.vmu.Unlock()
@@ -66,85 +66,80 @@ func (s *session) version() *version {
 	return s.stVersion
 }
 
-// Get current version; no barrier.
-func (s *session) version_NB() *version {
-	return s.stVersion
-}
-
 // Set current version to v.
 func (s *session) setVersion(v *version) {
 	s.vmu.Lock()
-	v.ref = 1
+	v.ref = 1 // Holds by session.
 	if old := s.stVersion; old != nil {
-		v.ref++
+		v.ref++ // Holds by old version.
 		old.next = v
-		old.release_NB()
+		old.releaseNB()
 	}
 	s.stVersion = v
 	s.vmu.Unlock()
 }
 
 // Get current unused file number.
-func (s *session) fileNum() uint64 {
-	return atomic.LoadUint64(&s.stFileNum)
+func (s *session) nextFileNum() uint64 {
+	return atomic.LoadUint64(&s.stNextFileNum)
 }
 
-// Get current unused file number to num.
-func (s *session) setFileNum(num uint64) {
-	atomic.StoreUint64(&s.stFileNum, num)
+// Set current unused file number to num.
+func (s *session) setNextFileNum(num uint64) {
+	atomic.StoreUint64(&s.stNextFileNum, num)
 }
 
 // Mark file number as used.
 func (s *session) markFileNum(num uint64) {
-	num += 1
+	nextFileNum := num + 1
 	for {
-		old, x := s.stFileNum, num
+		old, x := s.stNextFileNum, nextFileNum
 		if old > x {
 			x = old
 		}
-		if atomic.CompareAndSwapUint64(&s.stFileNum, old, x) {
+		if atomic.CompareAndSwapUint64(&s.stNextFileNum, old, x) {
 			break
 		}
 	}
 }
 
 // Allocate a file number.
-func (s *session) allocFileNum() (num uint64) {
-	return atomic.AddUint64(&s.stFileNum, 1) - 1
+func (s *session) allocFileNum() uint64 {
+	return atomic.AddUint64(&s.stNextFileNum, 1) - 1
 }
 
 // Reuse given file number.
 func (s *session) reuseFileNum(num uint64) {
 	for {
-		old, x := s.stFileNum, num
+		old, x := s.stNextFileNum, num
 		if old != x+1 {
 			x = old
 		}
-		if atomic.CompareAndSwapUint64(&s.stFileNum, old, x) {
+		if atomic.CompareAndSwapUint64(&s.stNextFileNum, old, x) {
 			break
 		}
 	}
 }
 
-// manifest related utils
+// Manifest related utils.
 
 // Fill given session record obj with current states; need external
 // synchronization.
 func (s *session) fillRecord(r *sessionRecord, snapshot bool) {
-	r.setNextNum(s.fileNum())
+	r.setNextFileNum(s.nextFileNum())
 
 	if snapshot {
 		if !r.has(recJournalNum) {
 			r.setJournalNum(s.stJournalNum)
 		}
 
-		if !r.has(recSeq) {
-			r.setSeq(s.stSeq)
+		if !r.has(recSeqNum) {
+			r.setSeqNum(s.stSeqNum)
 		}
 
-		for level, ik := range s.stCPtrs {
+		for level, ik := range s.stCompPtrs {
 			if ik != nil {
-				r.addCompactionPointer(level, ik)
+				r.addCompPtr(level, ik)
 			}
 		}
 
@@ -152,7 +147,7 @@ func (s *session) fillRecord(r *sessionRecord, snapshot bool) {
 	}
 }
 
-// Mark if record has been commited, this will update session state;
+// Mark if record has been committed, this will update session state;
 // need external synchronization.
 func (s *session) recordCommited(r *sessionRecord) {
 	if r.has(recJournalNum) {
@@ -163,12 +158,12 @@ func (s *session) recordCommited(r *sessionRecord) {
 		s.stPrevJournalNum = r.prevJournalNum
 	}
 
-	if r.has(recSeq) {
-		s.stSeq = r.seq
+	if r.has(recSeqNum) {
+		s.stSeqNum = r.seqNum
 	}
 
-	for _, p := range r.compactionPointers {
-		s.stCPtrs[p.level] = iKey(p.key)
+	for _, p := range r.compPtrs {
+		s.stCompPtrs[p.level] = iKey(p.ikey)
 	}
 }
 
@@ -183,10 +178,11 @@ func (s *session) newManifest(rec *sessionRecord, v *version) (err error) {
 	jw := journal.NewWriter(writer)
 
 	if v == nil {
-		v = s.version_NB()
+		v = s.version()
+		defer v.release()
 	}
 	if rec == nil {
-		rec = new(sessionRecord)
+		rec = &sessionRecord{numLevel: s.o.GetNumLevel()}
 	}
 	s.fillRecord(rec, true)
 	v.fillRecord(rec)
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage.go
index 75439f6db..46cc9d070 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage.go
@@ -221,7 +221,7 @@ func (fs *fileStorage) GetManifest() (f File, err error) {
 					fs.log(fmt.Sprintf("skipping %s: invalid file name", fn))
 					continue
 				}
-				if _, e1 := strconv.ParseUint(fn[7:], 10, 0); e1 != nil {
+				if _, e1 := strconv.ParseUint(fn[8:], 10, 0); e1 != nil {
 					fs.log(fmt.Sprintf("skipping %s: invalid file num: %v", fn, e1))
 					continue
 				}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_solaris.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_solaris.go
new file mode 100644
index 000000000..102031bfd
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_solaris.go
@@ -0,0 +1,68 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build solaris
+
+package storage
+
+import (
+	"os"
+	"syscall"
+)
+
+type unixFileLock struct {
+	f *os.File
+}
+
+func (fl *unixFileLock) release() error {
+	if err := setFileLock(fl.f, false); err != nil {
+		return err
+	}
+	return fl.f.Close()
+}
+
+func newFileLock(path string) (fl fileLock, err error) {
+	f, err := os.OpenFile(path, os.O_RDWR|os.O_CREATE, 0644)
+	if err != nil {
+		return
+	}
+	err = setFileLock(f, true)
+	if err != nil {
+		f.Close()
+		return
+	}
+	fl = &unixFileLock{f: f}
+	return
+}
+
+func setFileLock(f *os.File, lock bool) error {
+	flock := syscall.Flock_t{
+		Type:   syscall.F_UNLCK,
+		Start:  0,
+		Len:    0,
+		Whence: 1,
+	}
+	if lock {
+		flock.Type = syscall.F_WRLCK
+	}
+	return syscall.FcntlFlock(f.Fd(), syscall.F_SETLK, &flock)
+}
+
+func rename(oldpath, newpath string) error {
+	return os.Rename(oldpath, newpath)
+}
+
+func syncDir(name string) error {
+	f, err := os.Open(name)
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+	if err := f.Sync(); err != nil {
+		return err
+	}
+	return nil
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_unix.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_unix.go
index 73499afc8..d0a604b7a 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_unix.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/file_storage_unix.go
@@ -4,7 +4,7 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
-// +build darwin freebsd linux netbsd openbsd
+// +build darwin dragonfly freebsd linux netbsd openbsd
 
 package storage
 
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/storage.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/storage.go
index de5694888..85dd70b06 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/storage.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage/storage.go
@@ -67,7 +67,7 @@ type Writer interface {
 	Syncer
 }
 
-// File is the file.
+// File is the file. A file instance must be goroutine-safe.
 type File interface {
 	// Open opens the file for read. Returns os.ErrNotExist error
 	// if the file does not exist.
@@ -94,7 +94,7 @@ type File interface {
 	Remove() error
 }
 
-// Storage is the storage.
+// Storage is the storage. A storage instance must be goroutine-safe.
 type Storage interface {
 	// Lock locks the storage. Any subsequent attempt to call Lock will fail
 	// until the last lock released.
@@ -125,3 +125,33 @@ type Storage interface {
 	// Other methods should not be called after the storage has been closed.
 	Close() error
 }
+
+// FileInfo wraps basic file info.
+type FileInfo struct {
+	Type FileType
+	Num  uint64
+}
+
+func (fi FileInfo) String() string {
+	switch fi.Type {
+	case TypeManifest:
+		return fmt.Sprintf("MANIFEST-%06d", fi.Num)
+	case TypeJournal:
+		return fmt.Sprintf("%06d.log", fi.Num)
+	case TypeTable:
+		return fmt.Sprintf("%06d.ldb", fi.Num)
+	case TypeTemp:
+		return fmt.Sprintf("%06d.tmp", fi.Num)
+	default:
+		return fmt.Sprintf("%#x-%d", fi.Type, fi.Num)
+	}
+}
+
+// NewFileInfo creates new FileInfo from the given File. It will returns nil
+// if File is nil.
+func NewFileInfo(f File) *FileInfo {
+	if f == nil {
+		return nil
+	}
+	return &FileInfo{f.Type(), f.Num()}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage_test.go
index 27e76d707..dc1f1fb54 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/storage_test.go
@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"io"
 	"io/ioutil"
+	"math/rand"
 	"os"
 	"path/filepath"
 	"sync"
@@ -28,11 +29,25 @@ var (
 )
 
 var (
-	tsFSEnv  = os.Getenv("GOLEVELDB_USEFS")
-	tsKeepFS = tsFSEnv == "2"
-	tsFS     = tsKeepFS || tsFSEnv == "" || tsFSEnv == "1"
-	tsMU     = &sync.Mutex{}
-	tsNum    = 0
+	tsFSEnv   = os.Getenv("GOLEVELDB_USEFS")
+	tsTempdir = os.Getenv("GOLEVELDB_TEMPDIR")
+	tsKeepFS  = tsFSEnv == "2"
+	tsFS      = tsKeepFS || tsFSEnv == "" || tsFSEnv == "1"
+	tsMU      = &sync.Mutex{}
+	tsNum     = 0
+)
+
+type tsOp uint
+
+const (
+	tsOpOpen tsOp = iota
+	tsOpCreate
+	tsOpRead
+	tsOpReadAt
+	tsOpWrite
+	tsOpSync
+
+	tsOpNum
 )
 
 type tsLock struct {
@@ -53,6 +68,9 @@ type tsReader struct {
 func (tr tsReader) Read(b []byte) (n int, err error) {
 	ts := tr.tf.ts
 	ts.countRead(tr.tf.Type())
+	if tr.tf.shouldErrLocked(tsOpRead) {
+		return 0, errors.New("leveldb.testStorage: emulated read error")
+	}
 	n, err = tr.Reader.Read(b)
 	if err != nil && err != io.EOF {
 		ts.t.Errorf("E: read error, num=%d type=%v n=%d: %v", tr.tf.Num(), tr.tf.Type(), n, err)
@@ -63,6 +81,9 @@ func (tr tsReader) Read(b []byte) (n int, err error) {
 func (tr tsReader) ReadAt(b []byte, off int64) (n int, err error) {
 	ts := tr.tf.ts
 	ts.countRead(tr.tf.Type())
+	if tr.tf.shouldErrLocked(tsOpReadAt) {
+		return 0, errors.New("leveldb.testStorage: emulated readAt error")
+	}
 	n, err = tr.Reader.ReadAt(b, off)
 	if err != nil && err != io.EOF {
 		ts.t.Errorf("E: readAt error, num=%d type=%v off=%d n=%d: %v", tr.tf.Num(), tr.tf.Type(), off, n, err)
@@ -82,15 +103,12 @@ type tsWriter struct {
 }
 
 func (tw tsWriter) Write(b []byte) (n int, err error) {
-	ts := tw.tf.ts
-	ts.mu.Lock()
-	defer ts.mu.Unlock()
-	if ts.emuWriteErr&tw.tf.Type() != 0 {
+	if tw.tf.shouldErrLocked(tsOpWrite) {
 		return 0, errors.New("leveldb.testStorage: emulated write error")
 	}
 	n, err = tw.Writer.Write(b)
 	if err != nil {
-		ts.t.Errorf("E: write error, num=%d type=%v n=%d: %v", tw.tf.Num(), tw.tf.Type(), n, err)
+		tw.tf.ts.t.Errorf("E: write error, num=%d type=%v n=%d: %v", tw.tf.Num(), tw.tf.Type(), n, err)
 	}
 	return
 }
@@ -98,23 +116,23 @@ func (tw tsWriter) Write(b []byte) (n int, err error) {
 func (tw tsWriter) Sync() (err error) {
 	ts := tw.tf.ts
 	ts.mu.Lock()
-	defer ts.mu.Unlock()
 	for ts.emuDelaySync&tw.tf.Type() != 0 {
 		ts.cond.Wait()
 	}
-	if ts.emuSyncErr&tw.tf.Type() != 0 {
+	ts.mu.Unlock()
+	if tw.tf.shouldErrLocked(tsOpSync) {
 		return errors.New("leveldb.testStorage: emulated sync error")
 	}
 	err = tw.Writer.Sync()
 	if err != nil {
-		ts.t.Errorf("E: sync error, num=%d type=%v: %v", tw.tf.Num(), tw.tf.Type(), err)
+		tw.tf.ts.t.Errorf("E: sync error, num=%d type=%v: %v", tw.tf.Num(), tw.tf.Type(), err)
 	}
 	return
 }
 
 func (tw tsWriter) Close() (err error) {
 	err = tw.Writer.Close()
-	tw.tf.close("reader", err)
+	tw.tf.close("writer", err)
 	return
 }
 
@@ -127,6 +145,16 @@ func (tf tsFile) x() uint64 {
 	return tf.Num()<<typeShift | uint64(tf.Type())
 }
 
+func (tf tsFile) shouldErr(op tsOp) bool {
+	return tf.ts.shouldErr(tf, op)
+}
+
+func (tf tsFile) shouldErrLocked(op tsOp) bool {
+	tf.ts.mu.Lock()
+	defer tf.ts.mu.Unlock()
+	return tf.shouldErr(op)
+}
+
 func (tf tsFile) checkOpen(m string) error {
 	ts := tf.ts
 	if writer, ok := ts.opens[tf.x()]; ok {
@@ -163,7 +191,7 @@ func (tf tsFile) Open() (r storage.Reader, err error) {
 	if err != nil {
 		return
 	}
-	if ts.emuOpenErr&tf.Type() != 0 {
+	if tf.shouldErr(tsOpOpen) {
 		err = errors.New("leveldb.testStorage: emulated open error")
 		return
 	}
@@ -190,7 +218,7 @@ func (tf tsFile) Create() (w storage.Writer, err error) {
 	if err != nil {
 		return
 	}
-	if ts.emuCreateErr&tf.Type() != 0 {
+	if tf.shouldErr(tsOpCreate) {
 		err = errors.New("leveldb.testStorage: emulated create error")
 		return
 	}
@@ -205,6 +233,23 @@ func (tf tsFile) Create() (w storage.Writer, err error) {
 	return
 }
 
+func (tf tsFile) Replace(newfile storage.File) (err error) {
+	ts := tf.ts
+	ts.mu.Lock()
+	defer ts.mu.Unlock()
+	err = tf.checkOpen("replace")
+	if err != nil {
+		return
+	}
+	err = tf.File.Replace(newfile.(tsFile).File)
+	if err != nil {
+		ts.t.Errorf("E: cannot replace file, num=%d type=%v: %v", tf.Num(), tf.Type(), err)
+	} else {
+		ts.t.Logf("I: file replace, num=%d type=%v", tf.Num(), tf.Type())
+	}
+	return
+}
+
 func (tf tsFile) Remove() (err error) {
 	ts := tf.ts
 	ts.mu.Lock()
@@ -231,51 +276,75 @@ type testStorage struct {
 	cond sync.Cond
 	// Open files, true=writer, false=reader
 	opens         map[uint64]bool
-	emuOpenErr    storage.FileType
-	emuCreateErr  storage.FileType
 	emuDelaySync  storage.FileType
-	emuWriteErr   storage.FileType
-	emuSyncErr    storage.FileType
 	ignoreOpenErr storage.FileType
 	readCnt       uint64
 	readCntEn     storage.FileType
+
+	emuErr         [tsOpNum]storage.FileType
+	emuErrOnce     [tsOpNum]storage.FileType
+	emuRandErr     [tsOpNum]storage.FileType
+	emuRandErrProb int
+	emuErrOnceMap  map[uint64]uint
+	emuRandRand    *rand.Rand
+}
+
+func (ts *testStorage) shouldErr(tf tsFile, op tsOp) bool {
+	if ts.emuErr[op]&tf.Type() != 0 {
+		return true
+	} else if ts.emuRandErr[op]&tf.Type() != 0 || ts.emuErrOnce[op]&tf.Type() != 0 {
+		sop := uint(1) << op
+		eop := ts.emuErrOnceMap[tf.x()]
+		if eop&sop == 0 && (ts.emuRandRand.Int()%ts.emuRandErrProb == 0 || ts.emuErrOnce[op]&tf.Type() != 0) {
+			ts.emuErrOnceMap[tf.x()] = eop | sop
+			ts.t.Logf("I: emulated error: file=%d type=%v op=%v", tf.Num(), tf.Type(), op)
+			return true
+		}
+	}
+	return false
 }
 
-func (ts *testStorage) SetOpenErr(t storage.FileType) {
+func (ts *testStorage) SetEmuErr(t storage.FileType, ops ...tsOp) {
 	ts.mu.Lock()
-	ts.emuOpenErr = t
+	for _, op := range ops {
+		ts.emuErr[op] = t
+	}
 	ts.mu.Unlock()
 }
 
-func (ts *testStorage) SetCreateErr(t storage.FileType) {
+func (ts *testStorage) SetEmuErrOnce(t storage.FileType, ops ...tsOp) {
 	ts.mu.Lock()
-	ts.emuCreateErr = t
+	for _, op := range ops {
+		ts.emuErrOnce[op] = t
+	}
 	ts.mu.Unlock()
 }
 
-func (ts *testStorage) DelaySync(t storage.FileType) {
+func (ts *testStorage) SetEmuRandErr(t storage.FileType, ops ...tsOp) {
 	ts.mu.Lock()
-	ts.emuDelaySync |= t
-	ts.cond.Broadcast()
+	for _, op := range ops {
+		ts.emuRandErr[op] = t
+	}
 	ts.mu.Unlock()
 }
 
-func (ts *testStorage) ReleaseSync(t storage.FileType) {
+func (ts *testStorage) SetEmuRandErrProb(prob int) {
 	ts.mu.Lock()
-	ts.emuDelaySync &= ^t
-	ts.cond.Broadcast()
+	ts.emuRandErrProb = prob
 	ts.mu.Unlock()
 }
 
-func (ts *testStorage) SetWriteErr(t storage.FileType) {
+func (ts *testStorage) DelaySync(t storage.FileType) {
 	ts.mu.Lock()
-	ts.emuWriteErr = t
+	ts.emuDelaySync |= t
+	ts.cond.Broadcast()
 	ts.mu.Unlock()
 }
 
-func (ts *testStorage) SetSyncErr(t storage.FileType) {
+func (ts *testStorage) ReleaseSync(t storage.FileType) {
 	ts.mu.Lock()
-	ts.emuSyncErr = t
+	ts.emuDelaySync &= ^t
+	ts.cond.Broadcast()
 	ts.mu.Unlock()
 }
 
@@ -413,7 +482,11 @@ func newTestStorage(t *testing.T) *testStorage {
 			num := tsNum
 			tsNum++
 			tsMU.Unlock()
-			path := filepath.Join(os.TempDir(), fmt.Sprintf("goleveldb-test%d0%d0%d", os.Getuid(), os.Getpid(), num))
+			tempdir := tsTempdir
+			if tempdir == "" {
+				tempdir = os.TempDir()
+			}
+			path := filepath.Join(tempdir, fmt.Sprintf("goleveldb-test%d0%d0%d", os.Getuid(), os.Getpid(), num))
 			if _, err := os.Stat(path); err != nil {
 				stor, err = storage.OpenFile(path)
 				if err != nil {
@@ -436,6 +509,10 @@ func newTestStorage(t *testing.T) *testStorage {
 						}
 						f.Close()
 					}
+					if t.Failed() {
+						t.Logf("testing failed, test DB preserved at %s", path)
+						return nil
+					}
 					if tsKeepFS {
 						return nil
 					}
@@ -449,10 +526,13 @@ func newTestStorage(t *testing.T) *testStorage {
 		stor = storage.NewMemStorage()
 	}
 	ts := &testStorage{
-		t:       t,
-		Storage: stor,
-		closeFn: closeFn,
-		opens:   make(map[uint64]bool),
+		t:              t,
+		Storage:        stor,
+		closeFn:        closeFn,
+		opens:          make(map[uint64]bool),
+		emuErrOnceMap:  make(map[uint64]uint),
+		emuRandErrProb: 0x999,
+		emuRandRand:    rand.New(rand.NewSource(0xfacedead)),
 	}
 	ts.cond.L = &ts.mu
 	return ts
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table.go
index fdd5d2bcf..3e8df6af5 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table.go
@@ -7,11 +7,11 @@
 package leveldb
 
 import (
+	"fmt"
 	"sort"
 	"sync/atomic"
 
 	"github.com/syndtr/goleveldb/leveldb/cache"
-	"github.com/syndtr/goleveldb/leveldb/comparer"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/opt"
 	"github.com/syndtr/goleveldb/leveldb/storage"
@@ -19,34 +19,41 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-// table file
+// tFile holds basic information about a table.
 type tFile struct {
-	file     storage.File
-	seekLeft int32
-	size     uint64
-	min, max iKey
+	file       storage.File
+	seekLeft   int32
+	size       uint64
+	imin, imax iKey
 }
 
-// test if key is after t
-func (t *tFile) isAfter(key []byte, ucmp comparer.BasicComparer) bool {
-	return key != nil && ucmp.Compare(key, t.max.ukey()) > 0
+// Returns true if given key is after largest key of this table.
+func (t *tFile) after(icmp *iComparer, ukey []byte) bool {
+	return ukey != nil && icmp.uCompare(ukey, t.imax.ukey()) > 0
 }
 
-// test if key is before t
-func (t *tFile) isBefore(key []byte, ucmp comparer.BasicComparer) bool {
-	return key != nil && ucmp.Compare(key, t.min.ukey()) < 0
+// Returns true if given key is before smallest key of this table.
+func (t *tFile) before(icmp *iComparer, ukey []byte) bool {
+	return ukey != nil && icmp.uCompare(ukey, t.imin.ukey()) < 0
 }
 
-func (t *tFile) incrSeek() int32 {
+// Returns true if given key range overlaps with this table key range.
+func (t *tFile) overlaps(icmp *iComparer, umin, umax []byte) bool {
+	return !t.after(icmp, umin) && !t.before(icmp, umax)
+}
+
+// Cosumes one seek and return current seeks left.
+func (t *tFile) consumeSeek() int32 {
 	return atomic.AddInt32(&t.seekLeft, -1)
 }
 
-func newTFile(file storage.File, size uint64, min, max iKey) *tFile {
+// Creates new tFile.
+func newTableFile(file storage.File, size uint64, imin, imax iKey) *tFile {
 	f := &tFile{
 		file: file,
 		size: size,
-		min:  min,
-		max:  max,
+		imin: imin,
+		imax: imax,
 	}
 
 	// We arrange to automatically compact this file after
@@ -70,33 +77,52 @@ func newTFile(file storage.File, size uint64, min, max iKey) *tFile {
 	return f
 }
 
-// table files
+// tFiles hold multiple tFile.
 type tFiles []*tFile
 
 func (tf tFiles) Len() int      { return len(tf) }
 func (tf tFiles) Swap(i, j int) { tf[i], tf[j] = tf[j], tf[i] }
 
+func (tf tFiles) nums() string {
+	x := "[ "
+	for i, f := range tf {
+		if i != 0 {
+			x += ", "
+		}
+		x += fmt.Sprint(f.file.Num())
+	}
+	x += " ]"
+	return x
+}
+
+// Returns true if i smallest key is less than j.
+// This used for sort by key in ascending order.
 func (tf tFiles) lessByKey(icmp *iComparer, i, j int) bool {
 	a, b := tf[i], tf[j]
-	n := icmp.Compare(a.min, b.min)
+	n := icmp.Compare(a.imin, b.imin)
 	if n == 0 {
 		return a.file.Num() < b.file.Num()
 	}
 	return n < 0
 }
 
+// Returns true if i file number is greater than j.
+// This used for sort by file number in descending order.
 func (tf tFiles) lessByNum(i, j int) bool {
 	return tf[i].file.Num() > tf[j].file.Num()
 }
 
+// Sorts tables by key in ascending order.
 func (tf tFiles) sortByKey(icmp *iComparer) {
 	sort.Sort(&tFilesSortByKey{tFiles: tf, icmp: icmp})
 }
 
+// Sorts tables by file number in descending order.
 func (tf tFiles) sortByNum() {
 	sort.Sort(&tFilesSortByNum{tFiles: tf})
 }
 
+// Returns sum of all tables size.
 func (tf tFiles) size() (sum uint64) {
 	for _, t := range tf {
 		sum += t.size
@@ -104,94 +130,107 @@ func (tf tFiles) size() (sum uint64) {
 	return sum
 }
 
-func (tf tFiles) searchMin(key iKey, icmp *iComparer) int {
+// Searches smallest index of tables whose its smallest
+// key is after or equal with given key.
+func (tf tFiles) searchMin(icmp *iComparer, ikey iKey) int {
 	return sort.Search(len(tf), func(i int) bool {
-		return icmp.Compare(tf[i].min, key) >= 0
+		return icmp.Compare(tf[i].imin, ikey) >= 0
 	})
 }
 
-func (tf tFiles) searchMax(key iKey, icmp *iComparer) int {
+// Searches smallest index of tables whose its largest
+// key is after or equal with given key.
+func (tf tFiles) searchMax(icmp *iComparer, ikey iKey) int {
 	return sort.Search(len(tf), func(i int) bool {
-		return icmp.Compare(tf[i].max, key) >= 0
+		return icmp.Compare(tf[i].imax, ikey) >= 0
 	})
 }
 
-func (tf tFiles) isOverlaps(min, max []byte, disjSorted bool, icmp *iComparer) bool {
-	if !disjSorted {
-		// Need to check against all files
+// Returns true if given key range overlaps with one or more
+// tables key range. If unsorted is true then binary search will not be used.
+func (tf tFiles) overlaps(icmp *iComparer, umin, umax []byte, unsorted bool) bool {
+	if unsorted {
+		// Check against all files.
 		for _, t := range tf {
-			if !t.isAfter(min, icmp.ucmp) && !t.isBefore(max, icmp.ucmp) {
+			if t.overlaps(icmp, umin, umax) {
 				return true
 			}
 		}
 		return false
 	}
 
-	var idx int
-	if len(min) > 0 {
-		// Find the earliest possible internal key for min
-		idx = tf.searchMax(newIKey(min, kMaxSeq, tSeek), icmp)
+	i := 0
+	if len(umin) > 0 {
+		// Find the earliest possible internal key for min.
+		i = tf.searchMax(icmp, newIkey(umin, kMaxSeq, ktSeek))
 	}
-
-	if idx >= len(tf) {
-		// beginning of range is after all files, so no overlap
+	if i >= len(tf) {
+		// Beginning of range is after all files, so no overlap.
 		return false
 	}
-	return !tf[idx].isBefore(max, icmp.ucmp)
+	return !tf[i].before(icmp, umax)
 }
 
-func (tf tFiles) getOverlaps(min, max []byte, r *tFiles, disjSorted bool, ucmp comparer.BasicComparer) {
+// Returns tables whose its key range overlaps with given key range.
+// Range will be expanded if ukey found hop across tables.
+// If overlapped is true then the search will be restarted if umax
+// expanded.
+// The dst content will be overwritten.
+func (tf tFiles) getOverlaps(dst tFiles, icmp *iComparer, umin, umax []byte, overlapped bool) tFiles {
+	dst = dst[:0]
 	for i := 0; i < len(tf); {
 		t := tf[i]
-		i++
-		if t.isAfter(min, ucmp) || t.isBefore(max, ucmp) {
-			continue
-		}
-
-		*r = append(*r, t)
-		if !disjSorted {
-			// Level-0 files may overlap each other.  So check if the newly
-			// added file has expanded the range.  If so, restart search.
-			if min != nil && ucmp.Compare(t.min.ukey(), min) < 0 {
-				min = t.min.ukey()
-				*r = nil
-				i = 0
-			} else if max != nil && ucmp.Compare(t.max.ukey(), max) > 0 {
-				max = t.max.ukey()
-				*r = nil
+		if t.overlaps(icmp, umin, umax) {
+			if umin != nil && icmp.uCompare(t.imin.ukey(), umin) < 0 {
+				umin = t.imin.ukey()
+				dst = dst[:0]
 				i = 0
+				continue
+			} else if umax != nil && icmp.uCompare(t.imax.ukey(), umax) > 0 {
+				umax = t.imax.ukey()
+				// Restart search if it is overlapped.
+				if overlapped {
+					dst = dst[:0]
+					i = 0
+					continue
+				}
 			}
+
+			dst = append(dst, t)
 		}
+		i++
 	}
 
-	return
+	return dst
 }
 
-func (tf tFiles) getRange(icmp *iComparer) (min, max iKey) {
+// Returns tables key range.
+func (tf tFiles) getRange(icmp *iComparer) (imin, imax iKey) {
 	for i, t := range tf {
 		if i == 0 {
-			min, max = t.min, t.max
+			imin, imax = t.imin, t.imax
 			continue
 		}
-		if icmp.Compare(t.min, min) < 0 {
-			min = t.min
+		if icmp.Compare(t.imin, imin) < 0 {
+			imin = t.imin
 		}
-		if icmp.Compare(t.max, max) > 0 {
-			max = t.max
+		if icmp.Compare(t.imax, imax) > 0 {
+			imax = t.imax
 		}
 	}
 
 	return
 }
 
+// Creates iterator index from tables.
 func (tf tFiles) newIndexIterator(tops *tOps, icmp *iComparer, slice *util.Range, ro *opt.ReadOptions) iterator.IteratorIndexer {
 	if slice != nil {
 		var start, limit int
 		if slice.Start != nil {
-			start = tf.searchMax(iKey(slice.Start), icmp)
+			start = tf.searchMax(icmp, iKey(slice.Start))
 		}
 		if slice.Limit != nil {
-			limit = tf.searchMin(iKey(slice.Limit), icmp)
+			limit = tf.searchMin(icmp, iKey(slice.Limit))
 		} else {
 			limit = tf.Len()
 		}
@@ -206,6 +245,7 @@ func (tf tFiles) newIndexIterator(tops *tOps, icmp *iComparer, slice *util.Range
 	})
 }
 
+// Tables iterator index.
 type tFilesArrayIndexer struct {
 	tFiles
 	tops  *tOps
@@ -215,7 +255,7 @@ type tFilesArrayIndexer struct {
 }
 
 func (a *tFilesArrayIndexer) Search(key []byte) int {
-	return a.searchMax(iKey(key), a.icmp)
+	return a.searchMax(a.icmp, iKey(key))
 }
 
 func (a *tFilesArrayIndexer) Get(i int) iterator.Iterator {
@@ -225,6 +265,7 @@ func (a *tFilesArrayIndexer) Get(i int) iterator.Iterator {
 	return a.tops.newIterator(a.tFiles[i], nil, a.ro)
 }
 
+// Helper type for sortByKey.
 type tFilesSortByKey struct {
 	tFiles
 	icmp *iComparer
@@ -234,6 +275,7 @@ func (x *tFilesSortByKey) Less(i, j int) bool {
 	return x.lessByKey(x.icmp, i, j)
 }
 
+// Helper type for sortByNum.
 type tFilesSortByNum struct {
 	tFiles
 }
@@ -242,19 +284,15 @@ func (x *tFilesSortByNum) Less(i, j int) bool {
 	return x.lessByNum(i, j)
 }
 
-// table operations
+// Table operations.
 type tOps struct {
-	s       *session
-	cache   cache.Cache
-	cacheNS cache.Namespace
-}
-
-func newTableOps(s *session, cacheCap int) *tOps {
-	c := cache.NewLRUCache(cacheCap)
-	ns := c.GetNamespace(0)
-	return &tOps{s, c, ns}
+	s      *session
+	cache  *cache.Cache
+	bcache *cache.Cache
+	bpool  *util.BufferPool
 }
 
+// Creates an empty table and returns table writer.
 func (t *tOps) create() (*tWriter, error) {
 	file := t.s.getTableFile(t.s.allocFileNum())
 	fw, err := file.Create()
@@ -265,14 +303,15 @@ func (t *tOps) create() (*tWriter, error) {
 		t:    t,
 		file: file,
 		w:    fw,
-		tw:   table.NewWriter(fw, t.s.o),
+		tw:   table.NewWriter(fw, t.s.o.Options),
 	}, nil
 }
 
+// Builds table from src iterator.
 func (t *tOps) createFrom(src iterator.Iterator) (f *tFile, n int, err error) {
 	w, err := t.create()
 	if err != nil {
-		return f, n, err
+		return
 	}
 
 	defer func() {
@@ -282,7 +321,7 @@ func (t *tOps) createFrom(src iterator.Iterator) (f *tFile, n int, err error) {
 	}()
 
 	for src.Next() {
-		err = w.add(src.Key(), src.Value())
+		err = w.append(src.Key(), src.Value())
 		if err != nil {
 			return
 		}
@@ -297,84 +336,132 @@ func (t *tOps) createFrom(src iterator.Iterator) (f *tFile, n int, err error) {
 	return
 }
 
-func (t *tOps) lookup(f *tFile) (c cache.Object, err error) {
+// Opens table. It returns a cache handle, which should
+// be released after use.
+func (t *tOps) open(f *tFile) (ch *cache.Handle, err error) {
 	num := f.file.Num()
-	c, ok := t.cacheNS.Get(num, func() (ok bool, value interface{}, charge int, fin cache.SetFin) {
+	ch = t.cache.Get(0, num, func() (size int, value cache.Value) {
 		var r storage.Reader
 		r, err = f.file.Open()
 		if err != nil {
-			return
+			return 0, nil
 		}
 
-		o := t.s.o
-
-		var cacheNS cache.Namespace
-		if bc := o.GetBlockCache(); bc != nil {
-			cacheNS = bc.GetNamespace(num)
+		var bcache *cache.CacheGetter
+		if t.bcache != nil {
+			bcache = &cache.CacheGetter{Cache: t.bcache, NS: num}
 		}
 
-		ok = true
-		value = table.NewReader(r, int64(f.size), cacheNS, o)
-		charge = 1
-		fin = func() {
+		var tr *table.Reader
+		tr, err = table.NewReader(r, int64(f.size), storage.NewFileInfo(f.file), bcache, t.bpool, t.s.o.Options)
+		if err != nil {
 			r.Close()
+			return 0, nil
 		}
-		return
+		return 1, tr
+
 	})
-	if !ok && err == nil {
+	if ch == nil && err == nil {
 		err = ErrClosed
 	}
 	return
 }
 
-func (t *tOps) get(f *tFile, key []byte, ro *opt.ReadOptions) (rkey, rvalue []byte, err error) {
-	c, err := t.lookup(f)
+// Finds key/value pair whose key is greater than or equal to the
+// given key.
+func (t *tOps) find(f *tFile, key []byte, ro *opt.ReadOptions) (rkey, rvalue []byte, err error) {
+	ch, err := t.open(f)
 	if err != nil {
 		return nil, nil, err
 	}
-	defer c.Release()
-	return c.Value().(*table.Reader).Find(key, ro)
+	defer ch.Release()
+	return ch.Value().(*table.Reader).Find(key, true, ro)
+}
+
+// Finds key that is greater than or equal to the given key.
+func (t *tOps) findKey(f *tFile, key []byte, ro *opt.ReadOptions) (rkey []byte, err error) {
+	ch, err := t.open(f)
+	if err != nil {
+		return nil, err
+	}
+	defer ch.Release()
+	return ch.Value().(*table.Reader).FindKey(key, true, ro)
 }
 
+// Returns approximate offset of the given key.
 func (t *tOps) offsetOf(f *tFile, key []byte) (offset uint64, err error) {
-	c, err := t.lookup(f)
+	ch, err := t.open(f)
 	if err != nil {
 		return
 	}
-	_offset, err := c.Value().(*table.Reader).OffsetOf(key)
-	offset = uint64(_offset)
-	c.Release()
-	return
+	defer ch.Release()
+	offset_, err := ch.Value().(*table.Reader).OffsetOf(key)
+	return uint64(offset_), err
 }
 
+// Creates an iterator from the given table.
 func (t *tOps) newIterator(f *tFile, slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
-	c, err := t.lookup(f)
+	ch, err := t.open(f)
 	if err != nil {
 		return iterator.NewEmptyIterator(err)
 	}
-	iter := c.Value().(*table.Reader).NewIterator(slice, ro)
-	iter.SetReleaser(c)
+	iter := ch.Value().(*table.Reader).NewIterator(slice, ro)
+	iter.SetReleaser(ch)
 	return iter
 }
 
+// Removes table from persistent storage. It waits until
+// no one use the the table.
 func (t *tOps) remove(f *tFile) {
 	num := f.file.Num()
-	t.cacheNS.Delete(num, func(exist bool) {
+	t.cache.Delete(0, num, func() {
 		if err := f.file.Remove(); err != nil {
 			t.s.logf("table@remove removing @%d %q", num, err)
 		} else {
 			t.s.logf("table@remove removed @%d", num)
 		}
-		if bc := t.s.o.GetBlockCache(); bc != nil {
-			bc.GetNamespace(num).Zap(false)
+		if t.bcache != nil {
+			t.bcache.EvictNS(num)
 		}
 	})
 }
 
+// Closes the table ops instance. It will close all tables,
+// regadless still used or not.
 func (t *tOps) close() {
-	t.cache.Zap(true)
+	t.bpool.Close()
+	t.cache.Close()
+	if t.bcache != nil {
+		t.bcache.Close()
+	}
+}
+
+// Creates new initialized table ops instance.
+func newTableOps(s *session) *tOps {
+	var (
+		cacher cache.Cacher
+		bcache *cache.Cache
+	)
+	if s.o.GetOpenFilesCacheCapacity() > 0 {
+		cacher = cache.NewLRU(s.o.GetOpenFilesCacheCapacity())
+	}
+	if !s.o.DisableBlockCache {
+		var bcacher cache.Cacher
+		if s.o.GetBlockCacheCapacity() > 0 {
+			bcacher = cache.NewLRU(s.o.GetBlockCacheCapacity())
+		}
+		bcache = cache.NewCache(bcacher)
+	}
+	return &tOps{
+		s:      s,
+		cache:  cache.NewCache(cacher),
+		bcache: bcache,
+		bpool:  util.NewBufferPool(s.o.GetBlockSize() + 5),
+	}
 }
 
+// tWriter wraps the table writer. It keep track of file descriptor
+// and added key range.
 type tWriter struct {
 	t *tOps
 
@@ -385,7 +472,8 @@ type tWriter struct {
 	first, last []byte
 }
 
-func (w *tWriter) add(key, value []byte) error {
+// Append key/value pair to the table.
+func (w *tWriter) append(key, value []byte) error {
 	if w.first == nil {
 		w.first = append([]byte{}, key...)
 	}
@@ -393,30 +481,39 @@ func (w *tWriter) add(key, value []byte) error {
 	return w.tw.Append(key, value)
 }
 
+// Returns true if the table is empty.
 func (w *tWriter) empty() bool {
 	return w.first == nil
 }
 
+// Closes the storage.Writer.
+func (w *tWriter) close() {
+	if w.w != nil {
+		w.w.Close()
+		w.w = nil
+	}
+}
+
+// Finalizes the table and returns table file.
 func (w *tWriter) finish() (f *tFile, err error) {
+	defer w.close()
 	err = w.tw.Close()
 	if err != nil {
 		return
 	}
 	err = w.w.Sync()
 	if err != nil {
-		w.w.Close()
 		return
 	}
-	w.w.Close()
-	f = newTFile(w.file, uint64(w.tw.BytesLen()), iKey(w.first), iKey(w.last))
+	f = newTableFile(w.file, uint64(w.tw.BytesLen()), iKey(w.first), iKey(w.last))
 	return
 }
 
+// Drops the table.
 func (w *tWriter) drop() {
-	w.w.Close()
+	w.close()
 	w.file.Remove()
 	w.t.s.reuseFileNum(w.file.Num())
-	w.w = nil
 	w.file = nil
 	w.tw = nil
 	w.first = nil
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/block_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/block_test.go
index ca598f4f5..00e6f9eea 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/block_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/block_test.go
@@ -19,13 +19,18 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-func (b *block) TestNewIterator(slice *util.Range) iterator.Iterator {
-	return b.newIterator(slice, false, nil)
+type blockTesting struct {
+	tr *Reader
+	b  *block
+}
+
+func (t *blockTesting) TestNewIterator(slice *util.Range) iterator.Iterator {
+	return t.tr.newBlockIter(t.b, nil, slice, false)
 }
 
 var _ = testutil.Defer(func() {
 	Describe("Block", func() {
-		Build := func(kv *testutil.KeyValue, restartInterval int) *block {
+		Build := func(kv *testutil.KeyValue, restartInterval int) *blockTesting {
 			// Building the block.
 			bw := &blockWriter{
 				restartInterval: restartInterval,
@@ -39,11 +44,13 @@ var _ = testutil.Defer(func() {
 			// Opening the block.
 			data := bw.buf.Bytes()
 			restartsLen := int(binary.LittleEndian.Uint32(data[len(data)-4:]))
-			return &block{
-				cmp:            comparer.DefaultComparer,
-				data:           data,
-				restartsLen:    restartsLen,
-				restartsOffset: len(data) - (restartsLen+1)*4,
+			return &blockTesting{
+				tr: &Reader{cmp: comparer.DefaultComparer},
+				b: &block{
+					data:           data,
+					restartsLen:    restartsLen,
+					restartsOffset: len(data) - (restartsLen+1)*4,
+				},
 			}
 		}
 
@@ -59,7 +66,7 @@ var _ = testutil.Defer(func() {
 						// Make block.
 						br := Build(kv, restartInterval)
 						// Do testing.
-						testutil.KeyValueTesting(nil, br, kv.Clone())
+						testutil.KeyValueTesting(nil, kv.Clone(), br, nil, nil)
 					}
 
 					Describe(Text(), Test)
@@ -102,11 +109,11 @@ var _ = testutil.Defer(func() {
 			for restartInterval := 1; restartInterval <= 5; restartInterval++ {
 				Describe(fmt.Sprintf("with restart interval of %d", restartInterval), func() {
 					// Make block.
-					br := Build(kv, restartInterval)
+					bt := Build(kv, restartInterval)
 
 					Test := func(r *util.Range) func(done Done) {
 						return func(done Done) {
-							iter := br.newIterator(r, false, nil)
+							iter := bt.TestNewIterator(r)
 							Expect(iter.Error()).ShouldNot(HaveOccurred())
 
 							t := testutil.IteratorTesting{
@@ -115,6 +122,7 @@ var _ = testutil.Defer(func() {
 							}
 
 							testutil.DoIteratorTesting(&t)
+							iter.Release()
 							done <- true
 						}
 					}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/reader.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/reader.go
index 8acb9f720..6f38e84b3 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/reader.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/reader.go
@@ -8,27 +8,41 @@ package table
 
 import (
 	"encoding/binary"
-	"errors"
 	"fmt"
 	"io"
 	"sort"
 	"strings"
+	"sync"
 
-	"code.google.com/p/snappy-go/snappy"
+	"github.com/syndtr/gosnappy/snappy"
 
 	"github.com/syndtr/goleveldb/leveldb/cache"
 	"github.com/syndtr/goleveldb/leveldb/comparer"
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/filter"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/storage"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
 var (
-	ErrNotFound     = util.ErrNotFound
-	ErrIterReleased = errors.New("leveldb/table: iterator released")
+	ErrNotFound       = errors.ErrNotFound
+	ErrReaderReleased = errors.New("leveldb/table: reader released")
+	ErrIterReleased   = errors.New("leveldb/table: iterator released")
 )
 
+type ErrCorrupted struct {
+	Pos    int64
+	Size   int64
+	Kind   string
+	Reason string
+}
+
+func (e *ErrCorrupted) Error() string {
+	return fmt.Sprintf("leveldb/table: corruption on %s (pos=%d): %s", e.Kind, e.Pos, e.Reason)
+}
+
 func max(x, y int) int {
 	if x > y {
 		return x
@@ -37,40 +51,33 @@ func max(x, y int) int {
 }
 
 type block struct {
-	cmp            comparer.BasicComparer
+	bpool          *util.BufferPool
+	bh             blockHandle
 	data           []byte
 	restartsLen    int
 	restartsOffset int
-	// Whether checksum is verified and valid.
-	checksum bool
 }
 
-func (b *block) seek(rstart, rlimit int, key []byte) (index, offset int, err error) {
-	n := b.restartsOffset
-	data := b.data
-	cmp := b.cmp
-
+func (b *block) seek(cmp comparer.Comparer, rstart, rlimit int, key []byte) (index, offset int, err error) {
 	index = sort.Search(b.restartsLen-rstart-(b.restartsLen-rlimit), func(i int) bool {
-		offset := int(binary.LittleEndian.Uint32(data[n+4*(rstart+i):]))
-		offset += 1                               // shared always zero, since this is a restart point
-		v1, n1 := binary.Uvarint(data[offset:])   // key length
-		_, n2 := binary.Uvarint(data[offset+n1:]) // value length
+		offset := int(binary.LittleEndian.Uint32(b.data[b.restartsOffset+4*(rstart+i):]))
+		offset += 1                                 // shared always zero, since this is a restart point
+		v1, n1 := binary.Uvarint(b.data[offset:])   // key length
+		_, n2 := binary.Uvarint(b.data[offset+n1:]) // value length
 		m := offset + n1 + n2
-		return cmp.Compare(data[m:m+int(v1)], key) > 0
+		return cmp.Compare(b.data[m:m+int(v1)], key) > 0
 	}) + rstart - 1
 	if index < rstart {
 		// The smallest key is greater-than key sought.
 		index = rstart
 	}
-	offset = int(binary.LittleEndian.Uint32(data[n+4*index:]))
+	offset = int(binary.LittleEndian.Uint32(b.data[b.restartsOffset+4*index:]))
 	return
 }
 
 func (b *block) restartIndex(rstart, rlimit, offset int) int {
-	n := b.restartsOffset
-	data := b.data
 	return sort.Search(b.restartsLen-rstart-(b.restartsLen-rlimit), func(i int) bool {
-		return int(binary.LittleEndian.Uint32(data[n+4*(rstart+i):])) > offset
+		return int(binary.LittleEndian.Uint32(b.data[b.restartsOffset+4*(rstart+i):])) > offset
 	}) + rstart - 1
 }
 
@@ -81,7 +88,7 @@ func (b *block) restartOffset(index int) int {
 func (b *block) entry(offset int) (key, value []byte, nShared, n int, err error) {
 	if offset >= b.restartsOffset {
 		if offset != b.restartsOffset {
-			err = errors.New("leveldb/table: Reader: BlockEntry: invalid block (block entries offset not aligned)")
+			err = &ErrCorrupted{Reason: "entries offset not aligned"}
 		}
 		return
 	}
@@ -91,7 +98,7 @@ func (b *block) entry(offset int) (key, value []byte, nShared, n int, err error)
 	m := n0 + n1 + n2
 	n = m + int(v1) + int(v2)
 	if n0 <= 0 || n1 <= 0 || n2 <= 0 || offset+n > b.restartsOffset {
-		err = errors.New("leveldb/table: Reader: invalid block (block entries corrupted)")
+		err = &ErrCorrupted{Reason: "entries corrupted"}
 		return
 	}
 	key = b.data[offset+m : offset+m+int(v1)]
@@ -100,43 +107,10 @@ func (b *block) entry(offset int) (key, value []byte, nShared, n int, err error)
 	return
 }
 
-func (b *block) newIterator(slice *util.Range, inclLimit bool, cache util.Releaser) *blockIter {
-	bi := &blockIter{
-		block: b,
-		cache: cache,
-		// Valid key should never be nil.
-		key:             make([]byte, 0),
-		dir:             dirSOI,
-		riStart:         0,
-		riLimit:         b.restartsLen,
-		offsetStart:     0,
-		offsetRealStart: 0,
-		offsetLimit:     b.restartsOffset,
-	}
-	if slice != nil {
-		if slice.Start != nil {
-			if bi.Seek(slice.Start) {
-				bi.riStart = b.restartIndex(bi.restartIndex, b.restartsLen, bi.prevOffset)
-				bi.offsetStart = b.restartOffset(bi.riStart)
-				bi.offsetRealStart = bi.prevOffset
-			} else {
-				bi.riStart = b.restartsLen
-				bi.offsetStart = b.restartsOffset
-				bi.offsetRealStart = b.restartsOffset
-			}
-		}
-		if slice.Limit != nil {
-			if bi.Seek(slice.Limit) && (!inclLimit || bi.Next()) {
-				bi.offsetLimit = bi.prevOffset
-				bi.riLimit = bi.restartIndex + 1
-			}
-		}
-		bi.reset()
-		if bi.offsetStart > bi.offsetLimit {
-			bi.sErr(errors.New("leveldb/table: Reader: invalid slice range"))
-		}
-	}
-	return bi
+func (b *block) Release() {
+	b.bpool.Put(b.data)
+	b.bpool = nil
+	b.data = nil
 }
 
 type dir int
@@ -150,10 +124,12 @@ const (
 )
 
 type blockIter struct {
-	block           *block
-	cache, releaser util.Releaser
-	key, value      []byte
-	offset          int
+	tr            *Reader
+	block         *block
+	blockReleaser util.Releaser
+	releaser      util.Releaser
+	key, value    []byte
+	offset        int
 	// Previous offset, only filled by Next.
 	prevOffset   int
 	prevNode     []int
@@ -250,7 +226,7 @@ func (i *blockIter) Seek(key []byte) bool {
 		return false
 	}
 
-	ri, offset, err := i.block.seek(i.riStart, i.riLimit, key)
+	ri, offset, err := i.block.seek(i.tr.cmp, i.riStart, i.riLimit, key)
 	if err != nil {
 		i.sErr(err)
 		return false
@@ -261,7 +237,7 @@ func (i *blockIter) Seek(key []byte) bool {
 		i.dir = dirForward
 	}
 	for i.Next() {
-		if i.block.cmp.Compare(i.key, key) >= 0 {
+		if i.tr.cmp.Compare(i.key, key) >= 0 {
 			return true
 		}
 	}
@@ -286,7 +262,7 @@ func (i *blockIter) Next() bool {
 	for i.offset < i.offsetRealStart {
 		key, value, nShared, n, err := i.block.entry(i.offset)
 		if err != nil {
-			i.sErr(err)
+			i.sErr(i.tr.fixErrCorruptedBH(i.block.bh, err))
 			return false
 		}
 		if n == 0 {
@@ -300,13 +276,13 @@ func (i *blockIter) Next() bool {
 	if i.offset >= i.offsetLimit {
 		i.dir = dirEOI
 		if i.offset != i.offsetLimit {
-			i.sErr(errors.New("leveldb/table: Reader: Next: invalid block (block entries offset not aligned)"))
+			i.sErr(i.tr.newErrCorruptedBH(i.block.bh, "entries offset not aligned"))
 		}
 		return false
 	}
 	key, value, nShared, n, err := i.block.entry(i.offset)
 	if err != nil {
-		i.sErr(err)
+		i.sErr(i.tr.fixErrCorruptedBH(i.block.bh, err))
 		return false
 	}
 	if n == 0 {
@@ -391,7 +367,7 @@ func (i *blockIter) Prev() bool {
 	for {
 		key, value, nShared, n, err := i.block.entry(offset)
 		if err != nil {
-			i.sErr(err)
+			i.sErr(i.tr.fixErrCorruptedBH(i.block.bh, err))
 			return false
 		}
 		if offset >= i.offsetRealStart {
@@ -410,7 +386,7 @@ func (i *blockIter) Prev() bool {
 		// Stop if target offset reached.
 		if offset >= i.offset {
 			if offset != i.offset {
-				i.sErr(errors.New("leveldb/table: Reader: Prev: invalid block (block entries offset not aligned)"))
+				i.sErr(i.tr.newErrCorruptedBH(i.block.bh, "entries offset not aligned"))
 				return false
 			}
 
@@ -437,25 +413,33 @@ func (i *blockIter) Value() []byte {
 }
 
 func (i *blockIter) Release() {
-	i.prevNode = nil
-	i.prevKeys = nil
-	i.key = nil
-	i.value = nil
-	i.dir = dirReleased
-	if i.cache != nil {
-		i.cache.Release()
-		i.cache = nil
-	}
-	if i.releaser != nil {
-		i.releaser.Release()
-		i.releaser = nil
+	if i.dir != dirReleased {
+		i.tr = nil
+		i.block = nil
+		i.prevNode = nil
+		i.prevKeys = nil
+		i.key = nil
+		i.value = nil
+		i.dir = dirReleased
+		if i.blockReleaser != nil {
+			i.blockReleaser.Release()
+			i.blockReleaser = nil
+		}
+		if i.releaser != nil {
+			i.releaser.Release()
+			i.releaser = nil
+		}
 	}
 }
 
 func (i *blockIter) SetReleaser(releaser util.Releaser) {
-	if i.dir > dirReleased {
-		i.releaser = releaser
+	if i.dir == dirReleased {
+		panic(util.ErrReleased)
 	}
+	if i.releaser != nil && releaser != nil {
+		panic(util.ErrHasReleaser)
+	}
+	i.releaser = releaser
 }
 
 func (i *blockIter) Valid() bool {
@@ -467,21 +451,21 @@ func (i *blockIter) Error() error {
 }
 
 type filterBlock struct {
-	filter     filter.Filter
+	bpool      *util.BufferPool
 	data       []byte
 	oOffset    int
 	baseLg     uint
 	filtersNum int
 }
 
-func (b *filterBlock) contains(offset uint64, key []byte) bool {
+func (b *filterBlock) contains(filter filter.Filter, offset uint64, key []byte) bool {
 	i := int(offset >> b.baseLg)
 	if i < b.filtersNum {
 		o := b.data[b.oOffset+i*4:]
 		n := int(binary.LittleEndian.Uint32(o))
 		m := int(binary.LittleEndian.Uint32(o[4:]))
 		if n < m && m <= b.oOffset {
-			return b.filter.Contains(b.data[n:m], key)
+			return filter.Contains(b.data[n:m], key)
 		} else if n == m {
 			return false
 		}
@@ -489,12 +473,17 @@ func (b *filterBlock) contains(offset uint64, key []byte) bool {
 	return true
 }
 
+func (b *filterBlock) Release() {
+	b.bpool.Put(b.data)
+	b.bpool = nil
+	b.data = nil
+}
+
 type indexIter struct {
-	blockIter
-	tableReader *Reader
-	slice       *util.Range
+	*blockIter
+	tr    *Reader
+	slice *util.Range
 	// Options
-	checksum  bool
 	fillCache bool
 }
 
@@ -505,95 +494,173 @@ func (i *indexIter) Get() iterator.Iterator {
 	}
 	dataBH, n := decodeBlockHandle(value)
 	if n == 0 {
-		return iterator.NewEmptyIterator(errors.New("leveldb/table: Reader: invalid table (bad data block handle)"))
+		return iterator.NewEmptyIterator(i.tr.newErrCorruptedBH(i.tr.indexBH, "bad data block handle"))
 	}
+
 	var slice *util.Range
 	if i.slice != nil && (i.blockIter.isFirst() || i.blockIter.isLast()) {
 		slice = i.slice
 	}
-	return i.tableReader.getDataIter(dataBH, slice, i.checksum, i.fillCache)
+	return i.tr.getDataIterErr(dataBH, slice, i.tr.verifyChecksum, i.fillCache)
 }
 
 // Reader is a table reader.
 type Reader struct {
+	mu     sync.RWMutex
+	fi     *storage.FileInfo
 	reader io.ReaderAt
-	cache  cache.Namespace
+	cache  *cache.CacheGetter
 	err    error
+	bpool  *util.BufferPool
 	// Options
-	cmp        comparer.Comparer
-	filter     filter.Filter
-	checksum   bool
-	strictIter bool
+	o              *opt.Options
+	cmp            comparer.Comparer
+	filter         filter.Filter
+	verifyChecksum bool
 
-	dataEnd     int64
-	indexBlock  *block
-	filterBlock *filterBlock
+	dataEnd                   int64
+	metaBH, indexBH, filterBH blockHandle
+	indexBlock                *block
+	filterBlock               *filterBlock
 }
 
-func verifyChecksum(data []byte) bool {
-	n := len(data) - 4
-	checksum0 := binary.LittleEndian.Uint32(data[n:])
-	checksum1 := util.NewCRC(data[:n]).Value()
-	return checksum0 == checksum1
+func (r *Reader) blockKind(bh blockHandle) string {
+	switch bh.offset {
+	case r.metaBH.offset:
+		return "meta-block"
+	case r.indexBH.offset:
+		return "index-block"
+	case r.filterBH.offset:
+		if r.filterBH.length > 0 {
+			return "filter-block"
+		}
+	}
+	return "data-block"
 }
 
-func (r *Reader) readRawBlock(bh blockHandle, checksum bool) ([]byte, error) {
-	data := make([]byte, bh.length+blockTrailerLen)
+func (r *Reader) newErrCorrupted(pos, size int64, kind, reason string) error {
+	return &errors.ErrCorrupted{File: r.fi, Err: &ErrCorrupted{Pos: pos, Size: size, Kind: kind, Reason: reason}}
+}
+
+func (r *Reader) newErrCorruptedBH(bh blockHandle, reason string) error {
+	return r.newErrCorrupted(int64(bh.offset), int64(bh.length), r.blockKind(bh), reason)
+}
+
+func (r *Reader) fixErrCorruptedBH(bh blockHandle, err error) error {
+	if cerr, ok := err.(*ErrCorrupted); ok {
+		cerr.Pos = int64(bh.offset)
+		cerr.Size = int64(bh.length)
+		cerr.Kind = r.blockKind(bh)
+		return &errors.ErrCorrupted{File: r.fi, Err: cerr}
+	}
+	return err
+}
+
+func (r *Reader) readRawBlock(bh blockHandle, verifyChecksum bool) ([]byte, error) {
+	data := r.bpool.Get(int(bh.length + blockTrailerLen))
 	if _, err := r.reader.ReadAt(data, int64(bh.offset)); err != nil && err != io.EOF {
 		return nil, err
 	}
-	if checksum || r.checksum {
-		if !verifyChecksum(data) {
-			return nil, errors.New("leveldb/table: Reader: invalid block (checksum mismatch)")
+
+	if verifyChecksum {
+		n := bh.length + 1
+		checksum0 := binary.LittleEndian.Uint32(data[n:])
+		checksum1 := util.NewCRC(data[:n]).Value()
+		if checksum0 != checksum1 {
+			r.bpool.Put(data)
+			return nil, r.newErrCorruptedBH(bh, fmt.Sprintf("checksum mismatch, want=%#x got=%#x", checksum0, checksum1))
 		}
 	}
+
 	switch data[bh.length] {
 	case blockTypeNoCompression:
 		data = data[:bh.length]
 	case blockTypeSnappyCompression:
-		var err error
-		data, err = snappy.Decode(nil, data[:bh.length])
+		decLen, err := snappy.DecodedLen(data[:bh.length])
 		if err != nil {
-			return nil, err
+			return nil, r.newErrCorruptedBH(bh, err.Error())
+		}
+		decData := r.bpool.Get(decLen)
+		decData, err = snappy.Decode(decData, data[:bh.length])
+		r.bpool.Put(data)
+		if err != nil {
+			r.bpool.Put(decData)
+			return nil, r.newErrCorruptedBH(bh, err.Error())
 		}
+		data = decData
 	default:
-		return nil, fmt.Errorf("leveldb/table: Reader: unknown block compression type: %d", data[bh.length])
+		r.bpool.Put(data)
+		return nil, r.newErrCorruptedBH(bh, fmt.Sprintf("unknown compression type %#x", data[bh.length]))
 	}
 	return data, nil
 }
 
-func (r *Reader) readBlock(bh blockHandle, checksum bool) (*block, error) {
-	data, err := r.readRawBlock(bh, checksum)
+func (r *Reader) readBlock(bh blockHandle, verifyChecksum bool) (*block, error) {
+	data, err := r.readRawBlock(bh, verifyChecksum)
 	if err != nil {
 		return nil, err
 	}
 	restartsLen := int(binary.LittleEndian.Uint32(data[len(data)-4:]))
 	b := &block{
-		cmp:            r.cmp,
+		bpool:          r.bpool,
+		bh:             bh,
 		data:           data,
 		restartsLen:    restartsLen,
 		restartsOffset: len(data) - (restartsLen+1)*4,
-		checksum:       checksum || r.checksum,
 	}
 	return b, nil
 }
 
-func (r *Reader) readFilterBlock(bh blockHandle, filter filter.Filter) (*filterBlock, error) {
+func (r *Reader) readBlockCached(bh blockHandle, verifyChecksum, fillCache bool) (*block, util.Releaser, error) {
+	if r.cache != nil {
+		var (
+			err error
+			ch  *cache.Handle
+		)
+		if fillCache {
+			ch = r.cache.Get(bh.offset, func() (size int, value cache.Value) {
+				var b *block
+				b, err = r.readBlock(bh, verifyChecksum)
+				if err != nil {
+					return 0, nil
+				}
+				return cap(b.data), b
+			})
+		} else {
+			ch = r.cache.Get(bh.offset, nil)
+		}
+		if ch != nil {
+			b, ok := ch.Value().(*block)
+			if !ok {
+				ch.Release()
+				return nil, nil, errors.New("leveldb/table: inconsistent block type")
+			}
+			return b, ch, err
+		} else if err != nil {
+			return nil, nil, err
+		}
+	}
+
+	b, err := r.readBlock(bh, verifyChecksum)
+	return b, b, err
+}
+
+func (r *Reader) readFilterBlock(bh blockHandle) (*filterBlock, error) {
 	data, err := r.readRawBlock(bh, true)
 	if err != nil {
 		return nil, err
 	}
 	n := len(data)
 	if n < 5 {
-		return nil, errors.New("leveldb/table: Reader: invalid filter block (too short)")
+		return nil, r.newErrCorruptedBH(bh, "too short")
 	}
 	m := n - 5
 	oOffset := int(binary.LittleEndian.Uint32(data[m:]))
 	if oOffset > m {
-		return nil, errors.New("leveldb/table: Reader: invalid filter block (invalid offset)")
+		return nil, r.newErrCorruptedBH(bh, "invalid data-offsets offset")
 	}
 	b := &filterBlock{
-		filter:     filter,
+		bpool:      r.bpool,
 		data:       data,
 		oOffset:    oOffset,
 		baseLg:     uint(data[n-1]),
@@ -602,44 +669,111 @@ func (r *Reader) readFilterBlock(bh blockHandle, filter filter.Filter) (*filterB
 	return b, nil
 }
 
-func (r *Reader) getDataIter(dataBH blockHandle, slice *util.Range, checksum, fillCache bool) iterator.Iterator {
+func (r *Reader) readFilterBlockCached(bh blockHandle, fillCache bool) (*filterBlock, util.Releaser, error) {
 	if r.cache != nil {
-		// Get/set block cache.
-		var err error
-		cache, ok := r.cache.Get(dataBH.offset, func() (ok bool, value interface{}, charge int, fin cache.SetFin) {
-			if !fillCache {
-				return
+		var (
+			err error
+			ch  *cache.Handle
+		)
+		if fillCache {
+			ch = r.cache.Get(bh.offset, func() (size int, value cache.Value) {
+				var b *filterBlock
+				b, err = r.readFilterBlock(bh)
+				if err != nil {
+					return 0, nil
+				}
+				return cap(b.data), b
+			})
+		} else {
+			ch = r.cache.Get(bh.offset, nil)
+		}
+		if ch != nil {
+			b, ok := ch.Value().(*filterBlock)
+			if !ok {
+				ch.Release()
+				return nil, nil, errors.New("leveldb/table: inconsistent block type")
 			}
-			var dataBlock *block
-			dataBlock, err = r.readBlock(dataBH, checksum)
-			if err == nil {
-				ok = true
-				value = dataBlock
-				charge = int(dataBH.length)
+			return b, ch, err
+		} else if err != nil {
+			return nil, nil, err
+		}
+	}
+
+	b, err := r.readFilterBlock(bh)
+	return b, b, err
+}
+
+func (r *Reader) getIndexBlock(fillCache bool) (b *block, rel util.Releaser, err error) {
+	if r.indexBlock == nil {
+		return r.readBlockCached(r.indexBH, true, fillCache)
+	}
+	return r.indexBlock, util.NoopReleaser{}, nil
+}
+
+func (r *Reader) getFilterBlock(fillCache bool) (*filterBlock, util.Releaser, error) {
+	if r.filterBlock == nil {
+		return r.readFilterBlockCached(r.filterBH, fillCache)
+	}
+	return r.filterBlock, util.NoopReleaser{}, nil
+}
+
+func (r *Reader) newBlockIter(b *block, bReleaser util.Releaser, slice *util.Range, inclLimit bool) *blockIter {
+	bi := &blockIter{
+		tr:            r,
+		block:         b,
+		blockReleaser: bReleaser,
+		// Valid key should never be nil.
+		key:             make([]byte, 0),
+		dir:             dirSOI,
+		riStart:         0,
+		riLimit:         b.restartsLen,
+		offsetStart:     0,
+		offsetRealStart: 0,
+		offsetLimit:     b.restartsOffset,
+	}
+	if slice != nil {
+		if slice.Start != nil {
+			if bi.Seek(slice.Start) {
+				bi.riStart = b.restartIndex(bi.restartIndex, b.restartsLen, bi.prevOffset)
+				bi.offsetStart = b.restartOffset(bi.riStart)
+				bi.offsetRealStart = bi.prevOffset
+			} else {
+				bi.riStart = b.restartsLen
+				bi.offsetStart = b.restartsOffset
+				bi.offsetRealStart = b.restartsOffset
 			}
-			return
-		})
-		if err != nil {
-			return iterator.NewEmptyIterator(err)
 		}
-		if ok {
-			dataBlock := cache.Value().(*block)
-			if !dataBlock.checksum && (r.checksum || checksum) {
-				if !verifyChecksum(dataBlock.data) {
-					return iterator.NewEmptyIterator(errors.New("leveldb/table: Reader: invalid block (checksum mismatch)"))
-				}
-				dataBlock.checksum = true
+		if slice.Limit != nil {
+			if bi.Seek(slice.Limit) && (!inclLimit || bi.Next()) {
+				bi.offsetLimit = bi.prevOffset
+				bi.riLimit = bi.restartIndex + 1
 			}
-			iter := dataBlock.newIterator(slice, false, cache)
-			return iter
+		}
+		bi.reset()
+		if bi.offsetStart > bi.offsetLimit {
+			bi.sErr(errors.New("leveldb/table: invalid slice range"))
 		}
 	}
-	dataBlock, err := r.readBlock(dataBH, checksum)
+	return bi
+}
+
+func (r *Reader) getDataIter(dataBH blockHandle, slice *util.Range, verifyChecksum, fillCache bool) iterator.Iterator {
+	b, rel, err := r.readBlockCached(dataBH, verifyChecksum, fillCache)
 	if err != nil {
 		return iterator.NewEmptyIterator(err)
 	}
-	iter := dataBlock.newIterator(slice, false, nil)
-	return iter
+	return r.newBlockIter(b, rel, slice, false)
+}
+
+func (r *Reader) getDataIterErr(dataBH blockHandle, slice *util.Range, verifyChecksum, fillCache bool) iterator.Iterator {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	if r.err != nil {
+		return iterator.NewEmptyIterator(r.err)
+	}
+
+	return r.getDataIter(dataBH, slice, verifyChecksum, fillCache)
 }
 
 // NewIterator creates an iterator from the table.
@@ -653,35 +787,44 @@ func (r *Reader) getDataIter(dataBH blockHandle, slice *util.Range, checksum, fi
 // when not used.
 //
 // Also read Iterator documentation of the leveldb/iterator package.
-
 func (r *Reader) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
 	if r.err != nil {
 		return iterator.NewEmptyIterator(r.err)
 	}
 
+	fillCache := !ro.GetDontFillCache()
+	indexBlock, rel, err := r.getIndexBlock(fillCache)
+	if err != nil {
+		return iterator.NewEmptyIterator(err)
+	}
 	index := &indexIter{
-		blockIter:   *r.indexBlock.newIterator(slice, true, nil),
-		tableReader: r,
-		slice:       slice,
-		checksum:    ro.GetStrict(opt.StrictBlockChecksum),
-		fillCache:   !ro.GetDontFillCache(),
+		blockIter: r.newBlockIter(indexBlock, rel, slice, true),
+		tr:        r,
+		slice:     slice,
+		fillCache: !ro.GetDontFillCache(),
 	}
-	return iterator.NewIndexedIterator(index, r.strictIter || ro.GetStrict(opt.StrictIterator), false)
+	return iterator.NewIndexedIterator(index, opt.GetStrict(r.o, ro, opt.StrictReader))
 }
 
-// Find finds key/value pair whose key is greater than or equal to the
-// given key. It returns ErrNotFound if the table doesn't contain
-// such pair.
-//
-// The caller should not modify the contents of the returned slice, but
-// it is safe to modify the contents of the argument after Find returns.
-func (r *Reader) Find(key []byte, ro *opt.ReadOptions) (rkey, value []byte, err error) {
+func (r *Reader) find(key []byte, filtered bool, ro *opt.ReadOptions, noValue bool) (rkey, value []byte, err error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
 	if r.err != nil {
 		err = r.err
 		return
 	}
 
-	index := r.indexBlock.newIterator(nil, true, nil)
+	indexBlock, rel, err := r.getIndexBlock(true)
+	if err != nil {
+		return
+	}
+	defer rel.Release()
+
+	index := r.newBlockIter(indexBlock, nil, nil, true)
 	defer index.Release()
 	if !index.Seek(key) {
 		err = index.Error()
@@ -692,14 +835,23 @@ func (r *Reader) Find(key []byte, ro *opt.ReadOptions) (rkey, value []byte, err
 	}
 	dataBH, n := decodeBlockHandle(index.Value())
 	if n == 0 {
-		err = errors.New("leveldb/table: Reader: invalid table (bad data block handle)")
+		r.err = r.newErrCorruptedBH(r.indexBH, "bad data block handle")
 		return
 	}
-	if r.filterBlock != nil && !r.filterBlock.contains(dataBH.offset, key) {
-		err = ErrNotFound
-		return
+	if filtered && r.filter != nil {
+		filterBlock, frel, ferr := r.getFilterBlock(true)
+		if ferr == nil {
+			if !filterBlock.contains(r.filter, dataBH.offset, key) {
+				frel.Release()
+				return nil, nil, ErrNotFound
+			}
+			frel.Release()
+		} else if !errors.IsCorrupted(ferr) {
+			err = ferr
+			return
+		}
 	}
-	data := r.getDataIter(dataBH, nil, ro.GetStrict(opt.StrictBlockChecksum), !ro.GetDontFillCache())
+	data := r.getDataIter(dataBH, nil, r.verifyChecksum, !ro.GetDontFillCache())
 	defer data.Release()
 	if !data.Seek(key) {
 		err = data.Error()
@@ -708,23 +860,64 @@ func (r *Reader) Find(key []byte, ro *opt.ReadOptions) (rkey, value []byte, err
 		}
 		return
 	}
+	// Don't use block buffer, no need to copy the buffer.
 	rkey = data.Key()
-	value = data.Value()
+	if !noValue {
+		if r.bpool == nil {
+			value = data.Value()
+		} else {
+			// Use block buffer, and since the buffer will be recycled, the buffer
+			// need to be copied.
+			value = append([]byte{}, data.Value()...)
+		}
+	}
+	return
+}
+
+// Find finds key/value pair whose key is greater than or equal to the
+// given key. It returns ErrNotFound if the table doesn't contain
+// such pair.
+// If filtered is true then the nearest 'block' will be checked against
+// 'filter data' (if present) and will immediately return ErrNotFound if
+// 'filter data' indicates that such pair doesn't exist.
+//
+// The caller may modify the contents of the returned slice as it is its
+// own copy.
+// It is safe to modify the contents of the argument after Find returns.
+func (r *Reader) Find(key []byte, filtered bool, ro *opt.ReadOptions) (rkey, value []byte, err error) {
+	return r.find(key, filtered, ro, false)
+}
+
+// Find finds key that is greater than or equal to the given key.
+// It returns ErrNotFound if the table doesn't contain such key.
+// If filtered is true then the nearest 'block' will be checked against
+// 'filter data' (if present) and will immediately return ErrNotFound if
+// 'filter data' indicates that such key doesn't exist.
+//
+// The caller may modify the contents of the returned slice as it is its
+// own copy.
+// It is safe to modify the contents of the argument after Find returns.
+func (r *Reader) FindKey(key []byte, filtered bool, ro *opt.ReadOptions) (rkey []byte, err error) {
+	rkey, _, err = r.find(key, filtered, ro, true)
 	return
 }
 
 // Get gets the value for the given key. It returns errors.ErrNotFound
 // if the table does not contain the key.
 //
-// The caller should not modify the contents of the returned slice, but
-// it is safe to modify the contents of the argument after Get returns.
+// The caller may modify the contents of the returned slice as it is its
+// own copy.
+// It is safe to modify the contents of the argument after Find returns.
 func (r *Reader) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
 	if r.err != nil {
 		err = r.err
 		return
 	}
 
-	rkey, value, err := r.Find(key, ro)
+	rkey, value, err := r.find(key, false, ro, false)
 	if err == nil && r.cmp.Compare(rkey, key) != 0 {
 		value = nil
 		err = ErrNotFound
@@ -736,17 +929,26 @@ func (r *Reader) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error)
 //
 // It is safe to modify the contents of the argument after Get returns.
 func (r *Reader) OffsetOf(key []byte) (offset int64, err error) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
 	if r.err != nil {
 		err = r.err
 		return
 	}
 
-	index := r.indexBlock.newIterator(nil, true, nil)
+	indexBlock, rel, err := r.readBlockCached(r.indexBH, true, true)
+	if err != nil {
+		return
+	}
+	defer rel.Release()
+
+	index := r.newBlockIter(indexBlock, nil, nil, true)
 	defer index.Release()
 	if index.Seek(key) {
 		dataBH, n := decodeBlockHandle(index.Value())
 		if n == 0 {
-			err = errors.New("leveldb/table: Reader: invalid table (bad data block handle)")
+			r.err = r.newErrCorruptedBH(r.indexBH, "bad data block handle")
 			return
 		}
 		offset = int64(dataBH.offset)
@@ -759,90 +961,147 @@ func (r *Reader) OffsetOf(key []byte) (offset int64, err error) {
 	return
 }
 
-// NewReader creates a new initialized table reader for the file.
-// The cache is optional and can be nil.
-func NewReader(f io.ReaderAt, size int64, cache cache.Namespace, o *opt.Options) *Reader {
-	r := &Reader{
-		reader:     f,
-		cache:      cache,
-		cmp:        o.GetComparer(),
-		checksum:   o.GetStrict(opt.StrictBlockChecksum),
-		strictIter: o.GetStrict(opt.StrictIterator),
+// Release implements util.Releaser.
+// It also close the file if it is an io.Closer.
+func (r *Reader) Release() {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	if closer, ok := r.reader.(io.Closer); ok {
+		closer.Close()
+	}
+	if r.indexBlock != nil {
+		r.indexBlock.Release()
+		r.indexBlock = nil
 	}
+	if r.filterBlock != nil {
+		r.filterBlock.Release()
+		r.filterBlock = nil
+	}
+	r.reader = nil
+	r.cache = nil
+	r.bpool = nil
+	r.err = ErrReaderReleased
+}
+
+// NewReader creates a new initialized table reader for the file.
+// The fi, cache and bpool is optional and can be nil.
+//
+// The returned table reader instance is goroutine-safe.
+func NewReader(f io.ReaderAt, size int64, fi *storage.FileInfo, cache *cache.CacheGetter, bpool *util.BufferPool, o *opt.Options) (*Reader, error) {
 	if f == nil {
-		r.err = errors.New("leveldb/table: Reader: nil file")
-		return r
+		return nil, errors.New("leveldb/table: nil file")
 	}
+
+	r := &Reader{
+		fi:             fi,
+		reader:         f,
+		cache:          cache,
+		bpool:          bpool,
+		o:              o,
+		cmp:            o.GetComparer(),
+		verifyChecksum: o.GetStrict(opt.StrictBlockChecksum),
+	}
+
 	if size < footerLen {
-		r.err = errors.New("leveldb/table: Reader: invalid table (file size is too small)")
-		return r
+		r.err = r.newErrCorrupted(0, size, "table", "too small")
+		return r, nil
 	}
+
+	footerPos := size - footerLen
 	var footer [footerLen]byte
-	if _, err := r.reader.ReadAt(footer[:], size-footerLen); err != nil && err != io.EOF {
-		r.err = fmt.Errorf("leveldb/table: Reader: invalid table (could not read footer): %v", err)
+	if _, err := r.reader.ReadAt(footer[:], footerPos); err != nil && err != io.EOF {
+		return nil, err
 	}
 	if string(footer[footerLen-len(magic):footerLen]) != magic {
-		r.err = errors.New("leveldb/table: Reader: invalid table (bad magic number)")
-		return r
+		r.err = r.newErrCorrupted(footerPos, footerLen, "table-footer", "bad magic number")
+		return r, nil
 	}
+
+	var n int
 	// Decode the metaindex block handle.
-	metaBH, n := decodeBlockHandle(footer[:])
+	r.metaBH, n = decodeBlockHandle(footer[:])
 	if n == 0 {
-		r.err = errors.New("leveldb/table: Reader: invalid table (bad metaindex block handle)")
-		return r
+		r.err = r.newErrCorrupted(footerPos, footerLen, "table-footer", "bad metaindex block handle")
+		return r, nil
 	}
+
 	// Decode the index block handle.
-	indexBH, n := decodeBlockHandle(footer[n:])
+	r.indexBH, n = decodeBlockHandle(footer[n:])
 	if n == 0 {
-		r.err = errors.New("leveldb/table: Reader: invalid table (bad index block handle)")
-		return r
-	}
-	// Read index block.
-	r.indexBlock, r.err = r.readBlock(indexBH, true)
-	if r.err != nil {
-		return r
+		r.err = r.newErrCorrupted(footerPos, footerLen, "table-footer", "bad index block handle")
+		return r, nil
 	}
+
 	// Read metaindex block.
-	metaBlock, err := r.readBlock(metaBH, true)
+	metaBlock, err := r.readBlock(r.metaBH, true)
 	if err != nil {
-		r.err = err
-		return r
+		if errors.IsCorrupted(err) {
+			r.err = err
+			return r, nil
+		} else {
+			return nil, err
+		}
 	}
+
 	// Set data end.
-	r.dataEnd = int64(metaBH.offset)
-	metaIter := metaBlock.newIterator(nil, false, nil)
+	r.dataEnd = int64(r.metaBH.offset)
+
+	// Read metaindex.
+	metaIter := r.newBlockIter(metaBlock, nil, nil, true)
 	for metaIter.Next() {
 		key := string(metaIter.Key())
 		if !strings.HasPrefix(key, "filter.") {
 			continue
 		}
 		fn := key[7:]
-		var filter filter.Filter
 		if f0 := o.GetFilter(); f0 != nil && f0.Name() == fn {
-			filter = f0
+			r.filter = f0
 		} else {
 			for _, f0 := range o.GetAltFilters() {
 				if f0.Name() == fn {
-					filter = f0
+					r.filter = f0
 					break
 				}
 			}
 		}
-		if filter != nil {
+		if r.filter != nil {
 			filterBH, n := decodeBlockHandle(metaIter.Value())
 			if n == 0 {
 				continue
 			}
+			r.filterBH = filterBH
 			// Update data end.
 			r.dataEnd = int64(filterBH.offset)
-			filterBlock, err := r.readFilterBlock(filterBH, filter)
-			if err != nil {
-				continue
-			}
-			r.filterBlock = filterBlock
 			break
 		}
 	}
 	metaIter.Release()
-	return r
+	metaBlock.Release()
+
+	// Cache index and filter block locally, since we don't have global cache.
+	if cache == nil {
+		r.indexBlock, err = r.readBlock(r.indexBH, true)
+		if err != nil {
+			if errors.IsCorrupted(err) {
+				r.err = err
+				return r, nil
+			} else {
+				return nil, err
+			}
+		}
+		if r.filter != nil {
+			r.filterBlock, err = r.readFilterBlock(r.filterBH)
+			if err != nil {
+				if !errors.IsCorrupted(err) {
+					return nil, err
+				}
+
+				// Don't use filter then.
+				r.filter = nil
+			}
+		}
+	}
+
+	return r, nil
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table.go
index c0ac70d9e..beacdc1f0 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table.go
@@ -133,9 +133,9 @@ Filter block trailer:
 
       +- 4-bytes -+
      /             \
-    +---------------+---------------+---------------+-------------------------+------------------+
-    |    offset 1   |      ....     |    offset n   | filter offset (4-bytes) | base Lg (1-byte) |
-    +-------------- +---------------+---------------+-------------------------+------------------+
+    +---------------+---------------+---------------+-------------------------------+------------------+
+    | data 1 offset |      ....     | data n offset | data-offsets offset (4-bytes) | base Lg (1-byte) |
+    +-------------- +---------------+---------------+-------------------------------+------------------+
 
 
 NOTE: All fixed-length integer are little-endian.
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_suite_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_suite_test.go
index bc9eb83cc..6465da6e3 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_suite_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_suite_test.go
@@ -3,15 +3,9 @@ package table
 import (
 	"testing"
 
-	. "github.com/onsi/ginkgo"
-	. "github.com/onsi/gomega"
-
 	"github.com/syndtr/goleveldb/leveldb/testutil"
 )
 
 func TestTable(t *testing.T) {
-	testutil.RunDefer()
-
-	RegisterFailHandler(Fail)
-	RunSpecs(t, "Table Suite")
+	testutil.RunSuite(t, "Table Suite")
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_test.go
index d7d3b2a4b..4b59b31f5 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/table_test.go
@@ -23,7 +23,7 @@ type tableWrapper struct {
 }
 
 func (t tableWrapper) TestFind(key []byte) (rkey, rvalue []byte, err error) {
-	return t.Reader.Find(key, nil)
+	return t.Reader.Find(key, false, nil)
 }
 
 func (t tableWrapper) TestGet(key []byte) (value []byte, err error) {
@@ -59,7 +59,8 @@ var _ = testutil.Defer(func() {
 			It("Should be able to approximate offset of a key correctly", func() {
 				Expect(err).ShouldNot(HaveOccurred())
 
-				tr := NewReader(bytes.NewReader(buf.Bytes()), int64(buf.Len()), nil, o)
+				tr, err := NewReader(bytes.NewReader(buf.Bytes()), int64(buf.Len()), nil, nil, nil, o)
+				Expect(err).ShouldNot(HaveOccurred())
 				CheckOffset := func(key string, expect, threshold int) {
 					offset, err := tr.OffsetOf([]byte(key))
 					Expect(err).ShouldNot(HaveOccurred())
@@ -95,7 +96,7 @@ var _ = testutil.Defer(func() {
 				tw.Close()
 
 				// Opening the table.
-				tr := NewReader(bytes.NewReader(buf.Bytes()), int64(buf.Len()), nil, o)
+				tr, _ := NewReader(bytes.NewReader(buf.Bytes()), int64(buf.Len()), nil, nil, nil, o)
 				return tableWrapper{tr}
 			}
 			Test := func(kv *testutil.KeyValue, body func(r *Reader)) func() {
@@ -104,14 +105,16 @@ var _ = testutil.Defer(func() {
 					if body != nil {
 						body(db.(tableWrapper).Reader)
 					}
-					testutil.KeyValueTesting(nil, db, *kv)
+					testutil.KeyValueTesting(nil, *kv, db, nil, nil)
 				}
 			}
 
-			testutil.AllKeyValueTesting(nil, Build)
+			testutil.AllKeyValueTesting(nil, Build, nil, nil)
 			Describe("with one key per block", Test(testutil.KeyValue_Generate(nil, 9, 1, 10, 512, 512), func(r *Reader) {
 				It("should have correct blocks number", func() {
-					Expect(r.indexBlock.restartsLen).Should(Equal(9))
+					indexBlock, err := r.readBlock(r.indexBH, true)
+					Expect(err).To(BeNil())
+					Expect(indexBlock.restartsLen).Should(Equal(9))
 				})
 			}))
 		})
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/writer.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/writer.go
index 4e19e93a9..274c95fad 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/writer.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/table/writer.go
@@ -12,7 +12,7 @@ import (
 	"fmt"
 	"io"
 
-	"code.google.com/p/snappy-go/snappy"
+	"github.com/syndtr/gosnappy/snappy"
 
 	"github.com/syndtr/goleveldb/leveldb/comparer"
 	"github.com/syndtr/goleveldb/leveldb/filter"
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/db.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/db.go
index 4b87b5ef6..ec3f177a1 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/db.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/db.go
@@ -12,6 +12,7 @@ import (
 
 	. "github.com/onsi/gomega"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/iterator"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
@@ -34,6 +35,10 @@ type Get interface {
 	TestGet(key []byte) (value []byte, err error)
 }
 
+type Has interface {
+	TestHas(key []byte) (ret bool, err error)
+}
+
 type NewIterator interface {
 	TestNewIterator(slice *util.Range) iterator.Iterator
 }
@@ -110,7 +115,7 @@ func (t *DBTesting) TestAllPresent() {
 
 func (t *DBTesting) TestDeletedKey(key []byte) {
 	_, err := t.DB.TestGet(key)
-	Expect(err).Should(Equal(util.ErrNotFound), "Get on deleted key %q, %s", key, t.text())
+	Expect(err).Should(Equal(errors.ErrNotFound), "Get on deleted key %q, %s", key, t.text())
 }
 
 func (t *DBTesting) TestAllDeleted() {
@@ -212,5 +217,6 @@ func DoDBTesting(t *DBTesting) {
 		}
 
 		DoIteratorTesting(&it)
+		iter.Release()
 	}
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/ginkgo.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/ginkgo.go
new file mode 100644
index 000000000..82f3d0e81
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/ginkgo.go
@@ -0,0 +1,21 @@
+package testutil
+
+import (
+	. "github.com/onsi/ginkgo"
+	. "github.com/onsi/gomega"
+)
+
+func RunSuite(t GinkgoTestingT, name string) {
+	RunDefer()
+
+	SynchronizedBeforeSuite(func() []byte {
+		RunDefer("setup")
+		return nil
+	}, func(data []byte) {})
+	SynchronizedAfterSuite(func() {
+		RunDefer("teardown")
+	}, func() {})
+
+	RegisterFailHandler(Fail)
+	RunSpecs(t, name)
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/kvtest.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/kvtest.go
index 4fc75b6f2..a0b58f0e7 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/kvtest.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/kvtest.go
@@ -13,16 +13,28 @@ import (
 	. "github.com/onsi/ginkgo"
 	. "github.com/onsi/gomega"
 
+	"github.com/syndtr/goleveldb/leveldb/errors"
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-func KeyValueTesting(rnd *rand.Rand, p DB, kv KeyValue) {
+func KeyValueTesting(rnd *rand.Rand, kv KeyValue, p DB, setup func(KeyValue) DB, teardown func(DB)) {
 	if rnd == nil {
 		rnd = NewRand()
 	}
 
-	if db, ok := p.(Find); ok {
-		It("Should find all keys with Find", func() {
+	if p == nil {
+		BeforeEach(func() {
+			p = setup(kv)
+		})
+		if teardown != nil {
+			AfterEach(func() {
+				teardown(p)
+			})
+		}
+	}
+
+	It("Should find all keys with Find", func() {
+		if db, ok := p.(Find); ok {
 			ShuffledIndex(nil, kv.Len(), 1, func(i int) {
 				key_, key, value := kv.IndexInexact(i)
 
@@ -38,9 +50,11 @@ func KeyValueTesting(rnd *rand.Rand, p DB, kv KeyValue) {
 				Expect(rkey).Should(Equal(key))
 				Expect(rvalue).Should(Equal(value), "Value for key %q (%q)", key_, key)
 			})
-		})
+		}
+	})
 
-		It("Should return error if the key is not present", func() {
+	It("Should return error if the key is not present", func() {
+		if db, ok := p.(Find); ok {
 			var key []byte
 			if kv.Len() > 0 {
 				key_, _ := kv.Index(kv.Len() - 1)
@@ -48,12 +62,12 @@ func KeyValueTesting(rnd *rand.Rand, p DB, kv KeyValue) {
 			}
 			rkey, _, err := db.TestFind(key)
 			Expect(err).Should(HaveOccurred(), "Find for key %q yield key %q", key, rkey)
-			Expect(err).Should(Equal(util.ErrNotFound))
-		})
-	}
+			Expect(err).Should(Equal(errors.ErrNotFound))
+		}
+	})
 
-	if db, ok := p.(Get); ok {
-		It("Should only find exact key with Get", func() {
+	It("Should only find exact key with Get", func() {
+		if db, ok := p.(Get); ok {
 			ShuffledIndex(nil, kv.Len(), 1, func(i int) {
 				key_, key, value := kv.IndexInexact(i)
 
@@ -66,14 +80,34 @@ func KeyValueTesting(rnd *rand.Rand, p DB, kv KeyValue) {
 				if len(key_) > 0 {
 					_, err = db.TestGet(key_)
 					Expect(err).Should(HaveOccurred(), "Error for key %q", key_)
-					Expect(err).Should(Equal(util.ErrNotFound))
+					Expect(err).Should(Equal(errors.ErrNotFound))
 				}
 			})
-		})
-	}
+		}
+	})
+
+	It("Should only find present key with Has", func() {
+		if db, ok := p.(Has); ok {
+			ShuffledIndex(nil, kv.Len(), 1, func(i int) {
+				key_, key, _ := kv.IndexInexact(i)
+
+				// Using exact key.
+				ret, err := db.TestHas(key)
+				Expect(err).ShouldNot(HaveOccurred(), "Error for key %q", key)
+				Expect(ret).Should(BeTrue(), "False for key %q", key)
 
-	if db, ok := p.(NewIterator); ok {
-		TestIter := func(r *util.Range, _kv KeyValue) {
+				// Using inexact key.
+				if len(key_) > 0 {
+					ret, err = db.TestHas(key_)
+					Expect(err).ShouldNot(HaveOccurred(), "Error for key %q", key_)
+					Expect(ret).ShouldNot(BeTrue(), "True for key %q", key)
+				}
+			})
+		}
+	})
+
+	TestIter := func(r *util.Range, _kv KeyValue) {
+		if db, ok := p.(NewIterator); ok {
 			iter := db.TestNewIterator(r)
 			Expect(iter.Error()).ShouldNot(HaveOccurred())
 
@@ -83,46 +117,62 @@ func KeyValueTesting(rnd *rand.Rand, p DB, kv KeyValue) {
 			}
 
 			DoIteratorTesting(&t)
+			iter.Release()
 		}
+	}
 
-		It("Should iterates and seeks correctly", func(done Done) {
-			TestIter(nil, kv.Clone())
-			done <- true
-		}, 3.0)
-
-		RandomIndex(rnd, kv.Len(), kv.Len(), func(i int) {
-			type slice struct {
-				r            *util.Range
-				start, limit int
-			}
+	It("Should iterates and seeks correctly", func(done Done) {
+		TestIter(nil, kv.Clone())
+		done <- true
+	}, 3.0)
 
-			key_, _, _ := kv.IndexInexact(i)
-			for _, x := range []slice{
-				{&util.Range{Start: key_, Limit: nil}, i, kv.Len()},
-				{&util.Range{Start: nil, Limit: key_}, 0, i},
-			} {
-				It(fmt.Sprintf("Should iterates and seeks correctly of a slice %d .. %d", x.start, x.limit), func(done Done) {
-					TestIter(x.r, kv.Slice(x.start, x.limit))
-					done <- true
-				}, 3.0)
-			}
-		})
+	RandomIndex(rnd, kv.Len(), Min(kv.Len(), 50), func(i int) {
+		type slice struct {
+			r            *util.Range
+			start, limit int
+		}
 
-		RandomRange(rnd, kv.Len(), kv.Len(), func(start, limit int) {
-			It(fmt.Sprintf("Should iterates and seeks correctly of a slice %d .. %d", start, limit), func(done Done) {
-				r := kv.Range(start, limit)
-				TestIter(&r, kv.Slice(start, limit))
+		key_, _, _ := kv.IndexInexact(i)
+		for _, x := range []slice{
+			{&util.Range{Start: key_, Limit: nil}, i, kv.Len()},
+			{&util.Range{Start: nil, Limit: key_}, 0, i},
+		} {
+			It(fmt.Sprintf("Should iterates and seeks correctly of a slice %d .. %d", x.start, x.limit), func(done Done) {
+				TestIter(x.r, kv.Slice(x.start, x.limit))
 				done <- true
 			}, 3.0)
-		})
-	}
+		}
+	})
+
+	RandomRange(rnd, kv.Len(), Min(kv.Len(), 50), func(start, limit int) {
+		It(fmt.Sprintf("Should iterates and seeks correctly of a slice %d .. %d", start, limit), func(done Done) {
+			r := kv.Range(start, limit)
+			TestIter(&r, kv.Slice(start, limit))
+			done <- true
+		}, 3.0)
+	})
 }
 
-func AllKeyValueTesting(rnd *rand.Rand, body func(kv KeyValue) DB) {
+func AllKeyValueTesting(rnd *rand.Rand, body, setup func(KeyValue) DB, teardown func(DB)) {
 	Test := func(kv *KeyValue) func() {
 		return func() {
-			db := body(*kv)
-			KeyValueTesting(rnd, db, *kv)
+			var p DB
+			if setup != nil {
+				Defer("setup", func() {
+					p = setup(*kv)
+				})
+			}
+			if teardown != nil {
+				Defer("teardown", func() {
+					teardown(p)
+				})
+			}
+			if body != nil {
+				p = body(*kv)
+			}
+			KeyValueTesting(rnd, *kv, p, func(KeyValue) DB {
+				return p
+			}, nil)
 		}
 	}
 
@@ -133,4 +183,5 @@ func AllKeyValueTesting(rnd *rand.Rand, body func(kv KeyValue) DB) {
 	Describe("with big value", Test(KeyValue_BigValue()))
 	Describe("with special key", Test(KeyValue_SpecialKey()))
 	Describe("with multiple key/value", Test(KeyValue_MultipleKeyValue()))
+	Describe("with generated key/value", Test(KeyValue_Generate(nil, 120, 1, 50, 10, 120)))
 }
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/storage.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/storage.go
index 0f8d77a73..59c496d54 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/storage.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/storage.go
@@ -397,6 +397,7 @@ func (s *Storage) logI(format string, args ...interface{}) {
 
 func (s *Storage) Log(str string) {
 	s.log(1, "Log: "+str)
+	s.Storage.Log(str)
 }
 
 func (s *Storage) Lock() (r util.Releaser, err error) {
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/util.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/util.go
index 38fe25d52..97c5294b1 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/util.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil/util.go
@@ -155,3 +155,17 @@ func RandomRange(rnd *rand.Rand, n, round int, fn func(start, limit int)) {
 	}
 	return
 }
+
+func Max(x, y int) int {
+	if x > y {
+		return x
+	}
+	return y
+}
+
+func Min(x, y int) int {
+	if x < y {
+		return x
+	}
+	return y
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil_test.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil_test.go
index c1402fda3..25bf2b29f 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil_test.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/testutil_test.go
@@ -34,6 +34,10 @@ func (t *testingDB) TestGet(key []byte) (value []byte, err error) {
 	return t.Get(key, t.ro)
 }
 
+func (t *testingDB) TestHas(key []byte) (ret bool, err error) {
+	return t.Has(key, t.ro)
+}
+
 func (t *testingDB) TestNewIterator(slice *util.Range) iterator.Iterator {
 	return t.NewIterator(slice, t.ro)
 }
@@ -48,6 +52,7 @@ func (t *testingDB) TestClose() {
 func newTestingDB(o *opt.Options, ro *opt.ReadOptions, wo *opt.WriteOptions) *testingDB {
 	stor := testutil.NewStorage()
 	db, err := Open(stor, o)
+	// FIXME: This may be called from outside It, which may cause panic.
 	Expect(err).NotTo(HaveOccurred())
 	return &testingDB{
 		DB:   db,
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util.go
index a43d2e460..1a5bf71a3 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util.go
@@ -14,10 +14,10 @@ import (
 )
 
 func shorten(str string) string {
-	if len(str) <= 4 {
+	if len(str) <= 8 {
 		return str
 	}
-	return str[:1] + ".." + str[len(str)-1:]
+	return str[:3] + ".." + str[len(str)-3:]
 }
 
 var bunits = [...]string{"", "Ki", "Mi", "Gi"}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/buffer_pool.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/buffer_pool.go
new file mode 100644
index 000000000..2b8453d75
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/buffer_pool.go
@@ -0,0 +1,238 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package util
+
+import (
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+type buffer struct {
+	b    []byte
+	miss int
+}
+
+// BufferPool is a 'buffer pool'.
+type BufferPool struct {
+	pool      [6]chan []byte
+	size      [5]uint32
+	sizeMiss  [5]uint32
+	sizeHalf  [5]uint32
+	baseline  [4]int
+	baseline0 int
+
+	mu     sync.RWMutex
+	closed bool
+	closeC chan struct{}
+
+	get     uint32
+	put     uint32
+	half    uint32
+	less    uint32
+	equal   uint32
+	greater uint32
+	miss    uint32
+}
+
+func (p *BufferPool) poolNum(n int) int {
+	if n <= p.baseline0 && n > p.baseline0/2 {
+		return 0
+	}
+	for i, x := range p.baseline {
+		if n <= x {
+			return i + 1
+		}
+	}
+	return len(p.baseline) + 1
+}
+
+// Get returns buffer with length of n.
+func (p *BufferPool) Get(n int) []byte {
+	if p == nil {
+		return make([]byte, n)
+	}
+
+	p.mu.RLock()
+	defer p.mu.RUnlock()
+
+	if p.closed {
+		return make([]byte, n)
+	}
+
+	atomic.AddUint32(&p.get, 1)
+
+	poolNum := p.poolNum(n)
+	pool := p.pool[poolNum]
+	if poolNum == 0 {
+		// Fast path.
+		select {
+		case b := <-pool:
+			switch {
+			case cap(b) > n:
+				if cap(b)-n >= n {
+					atomic.AddUint32(&p.half, 1)
+					select {
+					case pool <- b:
+					default:
+					}
+					return make([]byte, n)
+				} else {
+					atomic.AddUint32(&p.less, 1)
+					return b[:n]
+				}
+			case cap(b) == n:
+				atomic.AddUint32(&p.equal, 1)
+				return b[:n]
+			default:
+				atomic.AddUint32(&p.greater, 1)
+			}
+		default:
+			atomic.AddUint32(&p.miss, 1)
+		}
+
+		return make([]byte, n, p.baseline0)
+	} else {
+		sizePtr := &p.size[poolNum-1]
+
+		select {
+		case b := <-pool:
+			switch {
+			case cap(b) > n:
+				if cap(b)-n >= n {
+					atomic.AddUint32(&p.half, 1)
+					sizeHalfPtr := &p.sizeHalf[poolNum-1]
+					if atomic.AddUint32(sizeHalfPtr, 1) == 20 {
+						atomic.StoreUint32(sizePtr, uint32(cap(b)/2))
+						atomic.StoreUint32(sizeHalfPtr, 0)
+					} else {
+						select {
+						case pool <- b:
+						default:
+						}
+					}
+					return make([]byte, n)
+				} else {
+					atomic.AddUint32(&p.less, 1)
+					return b[:n]
+				}
+			case cap(b) == n:
+				atomic.AddUint32(&p.equal, 1)
+				return b[:n]
+			default:
+				atomic.AddUint32(&p.greater, 1)
+				if uint32(cap(b)) >= atomic.LoadUint32(sizePtr) {
+					select {
+					case pool <- b:
+					default:
+					}
+				}
+			}
+		default:
+			atomic.AddUint32(&p.miss, 1)
+		}
+
+		if size := atomic.LoadUint32(sizePtr); uint32(n) > size {
+			if size == 0 {
+				atomic.CompareAndSwapUint32(sizePtr, 0, uint32(n))
+			} else {
+				sizeMissPtr := &p.sizeMiss[poolNum-1]
+				if atomic.AddUint32(sizeMissPtr, 1) == 20 {
+					atomic.StoreUint32(sizePtr, uint32(n))
+					atomic.StoreUint32(sizeMissPtr, 0)
+				}
+			}
+			return make([]byte, n)
+		} else {
+			return make([]byte, n, size)
+		}
+	}
+}
+
+// Put adds given buffer to the pool.
+func (p *BufferPool) Put(b []byte) {
+	if p == nil {
+		return
+	}
+
+	p.mu.RLock()
+	defer p.mu.RUnlock()
+
+	if p.closed {
+		return
+	}
+
+	atomic.AddUint32(&p.put, 1)
+
+	pool := p.pool[p.poolNum(cap(b))]
+	select {
+	case pool <- b:
+	default:
+	}
+
+}
+
+func (p *BufferPool) Close() {
+	if p == nil {
+		return
+	}
+
+	p.mu.Lock()
+	if !p.closed {
+		p.closed = true
+		p.closeC <- struct{}{}
+	}
+	p.mu.Unlock()
+}
+
+func (p *BufferPool) String() string {
+	if p == nil {
+		return "<nil>"
+	}
+
+	return fmt.Sprintf("BufferPool{B·%d Z·%v Zm·%v Zh·%v G·%d P·%d H·%d <·%d =·%d >·%d M·%d}",
+		p.baseline0, p.size, p.sizeMiss, p.sizeHalf, p.get, p.put, p.half, p.less, p.equal, p.greater, p.miss)
+}
+
+func (p *BufferPool) drain() {
+	ticker := time.NewTicker(2 * time.Second)
+	for {
+		select {
+		case <-ticker.C:
+			for _, ch := range p.pool {
+				select {
+				case <-ch:
+				default:
+				}
+			}
+		case <-p.closeC:
+			close(p.closeC)
+			for _, ch := range p.pool {
+				close(ch)
+			}
+			return
+		}
+	}
+}
+
+// NewBufferPool creates a new initialized 'buffer pool'.
+func NewBufferPool(baseline int) *BufferPool {
+	if baseline <= 0 {
+		panic("baseline can't be <= 0")
+	}
+	p := &BufferPool{
+		baseline0: baseline,
+		baseline:  [...]int{baseline / 4, baseline / 2, baseline * 2, baseline * 4},
+		closeC:    make(chan struct{}, 1),
+	}
+	for i, cap := range []int{2, 2, 4, 4, 2, 1} {
+		p.pool[i] = make(chan []byte, cap)
+	}
+	go p.drain()
+	return p
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/pool.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/pool.go
new file mode 100644
index 000000000..1f7fdd41f
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/pool.go
@@ -0,0 +1,21 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build go1.3
+
+package util
+
+import (
+	"sync"
+)
+
+type Pool struct {
+	sync.Pool
+}
+
+func NewPool(cap int) *Pool {
+	return &Pool{}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/pool_legacy.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/pool_legacy.go
new file mode 100644
index 000000000..27b8d03be
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/pool_legacy.go
@@ -0,0 +1,33 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build !go1.3
+
+package util
+
+type Pool struct {
+	pool chan interface{}
+}
+
+func (p *Pool) Get() interface{} {
+	select {
+	case x := <-p.pool:
+		return x
+	default:
+		return nil
+	}
+}
+
+func (p *Pool) Put(x interface{}) {
+	select {
+	case p.pool <- x:
+	default:
+	}
+}
+
+func NewPool(cap int) *Pool {
+	return &Pool{pool: make(chan interface{}, cap)}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/range.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/range.go
index da0583123..85159583d 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/range.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/range.go
@@ -14,3 +14,19 @@ type Range struct {
 	// Limit of the key range, not include in the range.
 	Limit []byte
 }
+
+// BytesPrefix returns key range that satisfy the given prefix.
+// This only applicable for the standard 'bytes comparer'.
+func BytesPrefix(prefix []byte) *Range {
+	var limit []byte
+	for i := len(prefix) - 1; i >= 0; i-- {
+		c := prefix[i]
+		if c < 0xff {
+			limit = make([]byte, i+1)
+			copy(limit, prefix)
+			limit[i] = c + 1
+			break
+		}
+	}
+	return &Range{prefix, limit}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/util.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/util.go
index 229c7d41f..f35976865 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/util.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/util/util.go
@@ -12,7 +12,8 @@ import (
 )
 
 var (
-	ErrNotFound = errors.New("leveldb: not found")
+	ErrReleased    = errors.New("leveldb: resource already relesed")
+	ErrHasReleaser = errors.New("leveldb: releaser already defined")
 )
 
 // Releaser is the interface that wraps the basic Release method.
@@ -27,23 +28,46 @@ type ReleaseSetter interface {
 	// SetReleaser associates the given releaser to the resources. The
 	// releaser will be called once coresponding resources released.
 	// Calling SetReleaser with nil will clear the releaser.
+	//
+	// This will panic if a releaser already present or coresponding
+	// resource is already released. Releaser should be cleared first
+	// before assigned a new one.
 	SetReleaser(releaser Releaser)
 }
 
 // BasicReleaser provides basic implementation of Releaser and ReleaseSetter.
 type BasicReleaser struct {
 	releaser Releaser
+	released bool
+}
+
+// Released returns whether Release method already called.
+func (r *BasicReleaser) Released() bool {
+	return r.released
 }
 
 // Release implements Releaser.Release.
 func (r *BasicReleaser) Release() {
-	if r.releaser != nil {
-		r.releaser.Release()
-		r.releaser = nil
+	if !r.released {
+		if r.releaser != nil {
+			r.releaser.Release()
+			r.releaser = nil
+		}
+		r.released = true
 	}
 }
 
 // SetReleaser implements ReleaseSetter.SetReleaser.
 func (r *BasicReleaser) SetReleaser(releaser Releaser) {
+	if r.released {
+		panic(ErrReleased)
+	}
+	if r.releaser != nil && releaser != nil {
+		panic(ErrHasReleaser)
+	}
 	r.releaser = releaser
 }
+
+type NoopReleaser struct{}
+
+func (NoopReleaser) Release() {}
diff --git a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/version.go b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/version.go
index 4c54d6480..88a52f53e 100644
--- a/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/version.go
+++ b/Godeps/_workspace/src/github.com/syndtr/goleveldb/leveldb/version.go
@@ -7,7 +7,6 @@
 package leveldb
 
 import (
-	"errors"
 	"sync/atomic"
 	"unsafe"
 
@@ -16,19 +15,6 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
-var levelMaxSize [kNumLevels]float64
-
-func init() {
-	// Precompute max size of each level
-	for level := range levelMaxSize {
-		res := float64(10 * 1048576)
-		for n := level; n > 1; n-- {
-			res *= 10
-		}
-		levelMaxSize[level] = res
-	}
-}
-
 type tSet struct {
 	level int
 	table *tFile
@@ -37,21 +23,26 @@ type tSet struct {
 type version struct {
 	s *session
 
-	tables [kNumLevels]tFiles
+	tables []tFiles
 
 	// Level that should be compacted next and its compaction score.
-	// Score < 1 means compaction is not strictly needed.  These fields
-	// are initialized by ComputeCompaction()
+	// Score < 1 means compaction is not strictly needed. These fields
+	// are initialized by computeCompaction()
 	cLevel int
 	cScore float64
 
 	cSeek unsafe.Pointer
 
-	ref  int
+	ref int
+	// Succeeding version.
 	next *version
 }
 
-func (v *version) release_NB() {
+func newVersion(s *session) *version {
+	return &version{s: s, tables: make([]tFiles, s.o.GetNumLevel())}
+}
+
+func (v *version) releaseNB() {
 	v.ref--
 	if v.ref > 0 {
 		return
@@ -60,8 +51,6 @@ func (v *version) release_NB() {
 		panic("negative version ref")
 	}
 
-	s := v.s
-
 	tables := make(map[uint64]bool)
 	for _, tt := range v.next.tables {
 		for _, t := range tt {
@@ -74,145 +63,184 @@ func (v *version) release_NB() {
 		for _, t := range tt {
 			num := t.file.Num()
 			if _, ok := tables[num]; !ok {
-				s.tops.remove(t)
+				v.s.tops.remove(t)
 			}
 		}
 	}
 
-	v.next.release_NB()
+	v.next.releaseNB()
 	v.next = nil
 }
 
 func (v *version) release() {
 	v.s.vmu.Lock()
-	v.release_NB()
+	v.releaseNB()
 	v.s.vmu.Unlock()
 }
 
-func (v *version) get(key iKey, ro *opt.ReadOptions) (value []byte, cstate bool, err error) {
-	s := v.s
-
-	ukey := key.ukey()
+func (v *version) walkOverlapping(ikey iKey, f func(level int, t *tFile) bool, lf func(level int) bool) {
+	ukey := ikey.ukey()
 
-	var tset *tSet
-	tseek := true
-
-	// We can search level-by-level since entries never hop across
-	// levels. Therefore we are guaranteed that if we find data
-	// in an smaller level, later levels are irrelevant.
-	for level, ts := range v.tables {
-		if len(ts) == 0 {
+	// Walk tables level-by-level.
+	for level, tables := range v.tables {
+		if len(tables) == 0 {
 			continue
 		}
 
 		if level == 0 {
 			// Level-0 files may overlap each other. Find all files that
-			// overlap user_key and process them in order from newest to
-			var tmp tFiles
-			for _, t := range ts {
-				if s.icmp.uCompare(ukey, t.min.ukey()) >= 0 &&
-					s.icmp.uCompare(ukey, t.max.ukey()) <= 0 {
-					tmp = append(tmp, t)
+			// overlap ukey.
+			for _, t := range tables {
+				if t.overlaps(v.s.icmp, ukey, ukey) {
+					if !f(level, t) {
+						return
+					}
 				}
 			}
-
-			if len(tmp) == 0 {
-				continue
-			}
-
-			tmp.sortByNum()
-			ts = tmp
 		} else {
-			i := ts.searchMax(key, s.icmp)
-			if i >= len(ts) || s.icmp.uCompare(ukey, ts[i].min.ukey()) < 0 {
-				continue
+			if i := tables.searchMax(v.s.icmp, ikey); i < len(tables) {
+				t := tables[i]
+				if v.s.icmp.uCompare(ukey, t.imin.ukey()) >= 0 {
+					if !f(level, t) {
+						return
+					}
+				}
 			}
+		}
 
-			ts = ts[i : i+1]
+		if lf != nil && !lf(level) {
+			return
 		}
+	}
+}
 
-		var l0found bool
-		var l0seq uint64
-		var l0type vType
-		var l0value []byte
-		for _, t := range ts {
-			if tseek {
-				if tset == nil {
-					tset = &tSet{level, t}
-				} else if tset.table.incrSeek() <= 0 {
-					cstate = atomic.CompareAndSwapPointer(&v.cSeek, nil, unsafe.Pointer(tset))
-					tseek = false
-				}
-			}
+func (v *version) get(ikey iKey, ro *opt.ReadOptions, noValue bool) (value []byte, tcomp bool, err error) {
+	ukey := ikey.ukey()
 
-			var _rkey, rval []byte
-			_rkey, rval, err = s.tops.get(t, key, ro)
-			if err == ErrNotFound {
-				continue
-			} else if err != nil {
-				return
+	var (
+		tset  *tSet
+		tseek bool
+
+		// Level-0.
+		zfound bool
+		zseq   uint64
+		zkt    kType
+		zval   []byte
+	)
+
+	err = ErrNotFound
+
+	// Since entries never hope across level, finding key/value
+	// in smaller level make later levels irrelevant.
+	v.walkOverlapping(ikey, func(level int, t *tFile) bool {
+		if !tseek {
+			if tset == nil {
+				tset = &tSet{level, t}
+			} else {
+				tseek = true
 			}
+		}
 
-			rkey := iKey(_rkey)
-			if seq, t, ok := rkey.parseNum(); ok {
-				if s.icmp.uCompare(ukey, rkey.ukey()) == 0 {
-					if level == 0 {
-						if seq >= l0seq {
-							l0found = true
-							l0seq = seq
-							l0type = t
-							l0value = rval
-						}
-					} else {
-						switch t {
-						case tVal:
-							value = rval
-						case tDel:
-							err = ErrNotFound
-						default:
-							panic("invalid type")
-						}
-						return
+		var (
+			fikey, fval []byte
+			ferr        error
+		)
+		if noValue {
+			fikey, ferr = v.s.tops.findKey(t, ikey, ro)
+		} else {
+			fikey, fval, ferr = v.s.tops.find(t, ikey, ro)
+		}
+		switch ferr {
+		case nil:
+		case ErrNotFound:
+			return true
+		default:
+			err = ferr
+			return false
+		}
+
+		if fukey, fseq, fkt, fkerr := parseIkey(fikey); fkerr == nil {
+			if v.s.icmp.uCompare(ukey, fukey) == 0 {
+				if level == 0 {
+					if fseq >= zseq {
+						zfound = true
+						zseq = fseq
+						zkt = fkt
+						zval = fval
 					}
+				} else {
+					switch fkt {
+					case ktVal:
+						value = fval
+						err = nil
+					case ktDel:
+					default:
+						panic("leveldb: invalid iKey type")
+					}
+					return false
 				}
-			} else {
-				err = errors.New("leveldb: internal key corrupted")
-				return
 			}
+		} else {
+			err = fkerr
+			return false
 		}
-		if level == 0 && l0found {
-			switch l0type {
-			case tVal:
-				value = l0value
-			case tDel:
-				err = ErrNotFound
+
+		return true
+	}, func(level int) bool {
+		if zfound {
+			switch zkt {
+			case ktVal:
+				value = zval
+				err = nil
+			case ktDel:
 			default:
-				panic("invalid type")
+				panic("leveldb: invalid iKey type")
 			}
-			return
+			return false
 		}
+
+		return true
+	})
+
+	if tseek && tset.table.consumeSeek() <= 0 {
+		tcomp = atomic.CompareAndSwapPointer(&v.cSeek, nil, unsafe.Pointer(tset))
 	}
 
-	err = ErrNotFound
 	return
 }
 
-func (v *version) getIterators(slice *util.Range, ro *opt.ReadOptions) (its []iterator.Iterator) {
-	s := v.s
+func (v *version) sampleSeek(ikey iKey) (tcomp bool) {
+	var tset *tSet
 
+	v.walkOverlapping(ikey, func(level int, t *tFile) bool {
+		if tset == nil {
+			tset = &tSet{level, t}
+			return true
+		} else {
+			if tset.table.consumeSeek() <= 0 {
+				tcomp = atomic.CompareAndSwapPointer(&v.cSeek, nil, unsafe.Pointer(tset))
+			}
+			return false
+		}
+	}, nil)
+
+	return
+}
+
+func (v *version) getIterators(slice *util.Range, ro *opt.ReadOptions) (its []iterator.Iterator) {
 	// Merge all level zero files together since they may overlap
 	for _, t := range v.tables[0] {
-		it := s.tops.newIterator(t, slice, ro)
+		it := v.s.tops.newIterator(t, slice, ro)
 		its = append(its, it)
 	}
 
-	strict := s.o.GetStrict(opt.StrictIterator) || ro.GetStrict(opt.StrictIterator)
-	for _, tt := range v.tables[1:] {
-		if len(tt) == 0 {
+	strict := opt.GetStrict(v.s.o.Options, ro, opt.StrictReader)
+	for _, tables := range v.tables[1:] {
+		if len(tables) == 0 {
 			continue
 		}
 
-		it := iterator.NewIndexedIterator(tt.newIndexIterator(s.tops, s.icmp, slice, ro), strict, true)
+		it := iterator.NewIndexedIterator(tables.newIndexIterator(v.s.tops, v.s.icmp, slice, ro), strict)
 		its = append(its, it)
 	}
 
@@ -220,7 +248,7 @@ func (v *version) getIterators(slice *util.Range, ro *opt.ReadOptions) (its []it
 }
 
 func (v *version) newStaging() *versionStaging {
-	return &versionStaging{base: v}
+	return &versionStaging{base: v, tables: make([]tablesScratch, v.s.o.GetNumLevel())}
 }
 
 // Spawn a new version based on this version.
@@ -242,25 +270,25 @@ func (v *version) tLen(level int) int {
 	return len(v.tables[level])
 }
 
-func (v *version) offsetOf(key iKey) (n uint64, err error) {
-	for level, tt := range v.tables {
-		for _, t := range tt {
-			if v.s.icmp.Compare(t.max, key) <= 0 {
-				// Entire file is before "key", so just add the file size
+func (v *version) offsetOf(ikey iKey) (n uint64, err error) {
+	for level, tables := range v.tables {
+		for _, t := range tables {
+			if v.s.icmp.Compare(t.imax, ikey) <= 0 {
+				// Entire file is before "ikey", so just add the file size
 				n += t.size
-			} else if v.s.icmp.Compare(t.min, key) > 0 {
-				// Entire file is after "key", so ignore
+			} else if v.s.icmp.Compare(t.imin, ikey) > 0 {
+				// Entire file is after "ikey", so ignore
 				if level > 0 {
 					// Files other than level 0 are sorted by meta->min, so
 					// no further files in this level will contain data for
-					// "key".
+					// "ikey".
 					break
 				}
 			} else {
-				// "key" falls in the range for this table.  Add the
-				// approximate offset of "key" within the table.
+				// "ikey" falls in the range for this table. Add the
+				// approximate offset of "ikey" within the table.
 				var nn uint64
-				nn, err = v.s.tops.offsetOf(t, key)
+				nn, err = v.s.tops.offsetOf(t, ikey)
 				if err != nil {
 					return 0, err
 				}
@@ -272,15 +300,16 @@ func (v *version) offsetOf(key iKey) (n uint64, err error) {
 	return
 }
 
-func (v *version) pickLevel(min, max []byte) (level int) {
-	if !v.tables[0].isOverlaps(min, max, false, v.s.icmp) {
-		var r tFiles
-		for ; level < kMaxMemCompactLevel; level++ {
-			if v.tables[level+1].isOverlaps(min, max, true, v.s.icmp) {
+func (v *version) pickLevel(umin, umax []byte) (level int) {
+	if !v.tables[0].overlaps(v.s.icmp, umin, umax, true) {
+		var overlaps tFiles
+		maxLevel := v.s.o.GetMaxMemCompationLevel()
+		for ; level < maxLevel; level++ {
+			if v.tables[level+1].overlaps(v.s.icmp, umin, umax, false) {
 				break
 			}
-			v.tables[level+2].getOverlaps(min, max, &r, true, v.s.icmp.ucmp)
-			if r.size() > kMaxGrandParentOverlapBytes {
+			overlaps = v.tables[level+2].getOverlaps(overlaps, v.s.icmp, umin, umax, false)
+			if overlaps.size() > uint64(v.s.o.GetCompactionGPOverlaps(level)) {
 				break
 			}
 		}
@@ -294,7 +323,7 @@ func (v *version) computeCompaction() {
 	var bestLevel int = -1
 	var bestScore float64 = -1
 
-	for level, ff := range v.tables {
+	for level, tables := range v.tables {
 		var score float64
 		if level == 0 {
 			// We treat level-0 specially by bounding the number of files
@@ -308,9 +337,9 @@ func (v *version) computeCompaction() {
 			// file size is small (perhaps because of a small write-buffer
 			// setting, or very high compression ratios, or lots of
 			// overwrites/deletions).
-			score = float64(len(ff)) / kL0_CompactionTrigger
+			score = float64(len(tables)) / float64(v.s.o.GetCompactionL0Trigger())
 		} else {
-			score = float64(ff.size()) / levelMaxSize[level]
+			score = float64(tables.size()) / float64(v.s.o.GetCompactionTotalSize(level))
 		}
 
 		if score > bestScore {
@@ -327,66 +356,62 @@ func (v *version) needCompaction() bool {
 	return v.cScore >= 1 || atomic.LoadPointer(&v.cSeek) != nil
 }
 
+type tablesScratch struct {
+	added   map[uint64]atRecord
+	deleted map[uint64]struct{}
+}
+
 type versionStaging struct {
 	base   *version
-	tables [kNumLevels]struct {
-		added   map[uint64]ntRecord
-		deleted map[uint64]struct{}
-	}
+	tables []tablesScratch
 }
 
 func (p *versionStaging) commit(r *sessionRecord) {
-	btt := p.base.tables
-
-	// deleted tables
-	for _, tr := range r.deletedTables {
-		tm := &(p.tables[tr.level])
+	// Deleted tables.
+	for _, r := range r.deletedTables {
+		tm := &(p.tables[r.level])
 
-		bt := btt[tr.level]
-		if len(bt) > 0 {
+		if len(p.base.tables[r.level]) > 0 {
 			if tm.deleted == nil {
 				tm.deleted = make(map[uint64]struct{})
 			}
-			tm.deleted[tr.num] = struct{}{}
+			tm.deleted[r.num] = struct{}{}
 		}
 
 		if tm.added != nil {
-			delete(tm.added, tr.num)
+			delete(tm.added, r.num)
 		}
 	}
 
-	// new tables
-	for _, tr := range r.addedTables {
-		tm := &(p.tables[tr.level])
+	// New tables.
+	for _, r := range r.addedTables {
+		tm := &(p.tables[r.level])
 
 		if tm.added == nil {
-			tm.added = make(map[uint64]ntRecord)
+			tm.added = make(map[uint64]atRecord)
 		}
-		tm.added[tr.num] = tr
+		tm.added[r.num] = r
 
 		if tm.deleted != nil {
-			delete(tm.deleted, tr.num)
+			delete(tm.deleted, r.num)
 		}
 	}
 }
 
 func (p *versionStaging) finish() *version {
-	s := p.base.s
-	btt := p.base.tables
-
-	// build new version
-	nv := &version{s: s}
+	// Build new version.
+	nv := newVersion(p.base.s)
 	for level, tm := range p.tables {
-		bt := btt[level]
+		btables := p.base.tables[level]
 
-		n := len(bt) + len(tm.added) - len(tm.deleted)
+		n := len(btables) + len(tm.added) - len(tm.deleted)
 		if n < 0 {
 			n = 0
 		}
 		nt := make(tFiles, 0, n)
 
-		// base tables
-		for _, t := range bt {
+		// Base tables.
+		for _, t := range btables {
 			if _, ok := tm.deleted[t.file.Num()]; ok {
 				continue
 			}
@@ -396,17 +421,21 @@ func (p *versionStaging) finish() *version {
 			nt = append(nt, t)
 		}
 
-		// new tables
-		for _, tr := range tm.added {
-			nt = append(nt, tr.makeFile(s))
+		// New tables.
+		for _, r := range tm.added {
+			nt = append(nt, p.base.s.tableFileFromRecord(r))
 		}
 
-		// sort tables
-		nt.sortByKey(s.icmp)
+		// Sort tables.
+		if level == 0 {
+			nt.sortByNum()
+		} else {
+			nt.sortByKey(p.base.s.icmp)
+		}
 		nv.tables[level] = nt
 	}
 
-	// compute compaction score for new version
+	// Compute compaction score for new version.
 	nv.computeCompaction()
 
 	return nv
@@ -421,7 +450,7 @@ func (vr *versionReleaser) Release() {
 	v := vr.v
 	v.s.vmu.Lock()
 	if !vr.once {
-		v.release_NB()
+		v.releaseNB()
 		vr.once = true
 	}
 	v.s.vmu.Unlock()
diff --git a/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/decode.go b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/decode.go
new file mode 100644
index 000000000..552a17bfb
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/decode.go
@@ -0,0 +1,292 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+var (
+	// ErrCorrupt reports that the input is invalid.
+	ErrCorrupt = errors.New("snappy: corrupt input")
+	// ErrUnsupported reports that the input isn't supported.
+	ErrUnsupported = errors.New("snappy: unsupported input")
+)
+
+// DecodedLen returns the length of the decoded block.
+func DecodedLen(src []byte) (int, error) {
+	v, _, err := decodedLen(src)
+	return v, err
+}
+
+// decodedLen returns the length of the decoded block and the number of bytes
+// that the length header occupied.
+func decodedLen(src []byte) (blockLen, headerLen int, err error) {
+	v, n := binary.Uvarint(src)
+	if n == 0 {
+		return 0, 0, ErrCorrupt
+	}
+	if uint64(int(v)) != v {
+		return 0, 0, errors.New("snappy: decoded block is too large")
+	}
+	return int(v), n, nil
+}
+
+// Decode returns the decoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire decoded block.
+// Otherwise, a newly allocated slice will be returned.
+// It is valid to pass a nil dst.
+func Decode(dst, src []byte) ([]byte, error) {
+	dLen, s, err := decodedLen(src)
+	if err != nil {
+		return nil, err
+	}
+	if len(dst) < dLen {
+		dst = make([]byte, dLen)
+	}
+
+	var d, offset, length int
+	for s < len(src) {
+		switch src[s] & 0x03 {
+		case tagLiteral:
+			x := uint(src[s] >> 2)
+			switch {
+			case x < 60:
+				s += 1
+			case x == 60:
+				s += 2
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-1])
+			case x == 61:
+				s += 3
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-2]) | uint(src[s-1])<<8
+			case x == 62:
+				s += 4
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-3]) | uint(src[s-2])<<8 | uint(src[s-1])<<16
+			case x == 63:
+				s += 5
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-4]) | uint(src[s-3])<<8 | uint(src[s-2])<<16 | uint(src[s-1])<<24
+			}
+			length = int(x + 1)
+			if length <= 0 {
+				return nil, errors.New("snappy: unsupported literal length")
+			}
+			if length > len(dst)-d || length > len(src)-s {
+				return nil, ErrCorrupt
+			}
+			copy(dst[d:], src[s:s+length])
+			d += length
+			s += length
+			continue
+
+		case tagCopy1:
+			s += 2
+			if s > len(src) {
+				return nil, ErrCorrupt
+			}
+			length = 4 + int(src[s-2])>>2&0x7
+			offset = int(src[s-2])&0xe0<<3 | int(src[s-1])
+
+		case tagCopy2:
+			s += 3
+			if s > len(src) {
+				return nil, ErrCorrupt
+			}
+			length = 1 + int(src[s-3])>>2
+			offset = int(src[s-2]) | int(src[s-1])<<8
+
+		case tagCopy4:
+			return nil, errors.New("snappy: unsupported COPY_4 tag")
+		}
+
+		end := d + length
+		if offset > d || end > len(dst) {
+			return nil, ErrCorrupt
+		}
+		for ; d < end; d++ {
+			dst[d] = dst[d-offset]
+		}
+	}
+	if d != dLen {
+		return nil, ErrCorrupt
+	}
+	return dst[:d], nil
+}
+
+// NewReader returns a new Reader that decompresses from r, using the framing
+// format described at
+// https://code.google.com/p/snappy/source/browse/trunk/framing_format.txt
+func NewReader(r io.Reader) *Reader {
+	return &Reader{
+		r:       r,
+		decoded: make([]byte, maxUncompressedChunkLen),
+		buf:     make([]byte, MaxEncodedLen(maxUncompressedChunkLen)+checksumSize),
+	}
+}
+
+// Reader is an io.Reader than can read Snappy-compressed bytes.
+type Reader struct {
+	r       io.Reader
+	err     error
+	decoded []byte
+	buf     []byte
+	// decoded[i:j] contains decoded bytes that have not yet been passed on.
+	i, j       int
+	readHeader bool
+}
+
+// Reset discards any buffered data, resets all state, and switches the Snappy
+// reader to read from r. This permits reusing a Reader rather than allocating
+// a new one.
+func (r *Reader) Reset(reader io.Reader) {
+	r.r = reader
+	r.err = nil
+	r.i = 0
+	r.j = 0
+	r.readHeader = false
+}
+
+func (r *Reader) readFull(p []byte) (ok bool) {
+	if _, r.err = io.ReadFull(r.r, p); r.err != nil {
+		if r.err == io.ErrUnexpectedEOF {
+			r.err = ErrCorrupt
+		}
+		return false
+	}
+	return true
+}
+
+// Read satisfies the io.Reader interface.
+func (r *Reader) Read(p []byte) (int, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+	for {
+		if r.i < r.j {
+			n := copy(p, r.decoded[r.i:r.j])
+			r.i += n
+			return n, nil
+		}
+		if !r.readFull(r.buf[:4]) {
+			return 0, r.err
+		}
+		chunkType := r.buf[0]
+		if !r.readHeader {
+			if chunkType != chunkTypeStreamIdentifier {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			r.readHeader = true
+		}
+		chunkLen := int(r.buf[1]) | int(r.buf[2])<<8 | int(r.buf[3])<<16
+		if chunkLen > len(r.buf) {
+			r.err = ErrUnsupported
+			return 0, r.err
+		}
+
+		// The chunk types are specified at
+		// https://code.google.com/p/snappy/source/browse/trunk/framing_format.txt
+		switch chunkType {
+		case chunkTypeCompressedData:
+			// Section 4.2. Compressed data (chunk type 0x00).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			buf := r.buf[:chunkLen]
+			if !r.readFull(buf) {
+				return 0, r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			buf = buf[checksumSize:]
+
+			n, err := DecodedLen(buf)
+			if err != nil {
+				r.err = err
+				return 0, r.err
+			}
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			if _, err := Decode(r.decoded, buf); err != nil {
+				r.err = err
+				return 0, r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeUncompressedData:
+			// Section 4.3. Uncompressed data (chunk type 0x01).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			buf := r.buf[:checksumSize]
+			if !r.readFull(buf) {
+				return 0, r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			// Read directly into r.decoded instead of via r.buf.
+			n := chunkLen - checksumSize
+			if !r.readFull(r.decoded[:n]) {
+				return 0, r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeStreamIdentifier:
+			// Section 4.1. Stream identifier (chunk type 0xff).
+			if chunkLen != len(magicBody) {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			if !r.readFull(r.buf[:len(magicBody)]) {
+				return 0, r.err
+			}
+			for i := 0; i < len(magicBody); i++ {
+				if r.buf[i] != magicBody[i] {
+					r.err = ErrCorrupt
+					return 0, r.err
+				}
+			}
+			continue
+		}
+
+		if chunkType <= 0x7f {
+			// Section 4.5. Reserved unskippable chunks (chunk types 0x02-0x7f).
+			r.err = ErrUnsupported
+			return 0, r.err
+
+		} else {
+			// Section 4.4 Padding (chunk type 0xfe).
+			// Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd).
+			if !r.readFull(r.buf[:chunkLen]) {
+				return 0, r.err
+			}
+		}
+	}
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/encode.go b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/encode.go
new file mode 100644
index 000000000..dda372422
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/encode.go
@@ -0,0 +1,258 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"encoding/binary"
+	"io"
+)
+
+// We limit how far copy back-references can go, the same as the C++ code.
+const maxOffset = 1 << 15
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+func emitLiteral(dst, lit []byte) int {
+	i, n := 0, uint(len(lit)-1)
+	switch {
+	case n < 60:
+		dst[0] = uint8(n)<<2 | tagLiteral
+		i = 1
+	case n < 1<<8:
+		dst[0] = 60<<2 | tagLiteral
+		dst[1] = uint8(n)
+		i = 2
+	case n < 1<<16:
+		dst[0] = 61<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		i = 3
+	case n < 1<<24:
+		dst[0] = 62<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		dst[3] = uint8(n >> 16)
+		i = 4
+	case int64(n) < 1<<32:
+		dst[0] = 63<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		dst[3] = uint8(n >> 16)
+		dst[4] = uint8(n >> 24)
+		i = 5
+	default:
+		panic("snappy: source buffer is too long")
+	}
+	if copy(dst[i:], lit) != len(lit) {
+		panic("snappy: destination buffer is too short")
+	}
+	return i + len(lit)
+}
+
+// emitCopy writes a copy chunk and returns the number of bytes written.
+func emitCopy(dst []byte, offset, length int) int {
+	i := 0
+	for length > 0 {
+		x := length - 4
+		if 0 <= x && x < 1<<3 && offset < 1<<11 {
+			dst[i+0] = uint8(offset>>8)&0x07<<5 | uint8(x)<<2 | tagCopy1
+			dst[i+1] = uint8(offset)
+			i += 2
+			break
+		}
+
+		x = length
+		if x > 1<<6 {
+			x = 1 << 6
+		}
+		dst[i+0] = uint8(x-1)<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= x
+	}
+	return i
+}
+
+// Encode returns the encoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire encoded block.
+// Otherwise, a newly allocated slice will be returned.
+// It is valid to pass a nil dst.
+func Encode(dst, src []byte) ([]byte, error) {
+	if n := MaxEncodedLen(len(src)); len(dst) < n {
+		dst = make([]byte, n)
+	}
+
+	// The block starts with the varint-encoded length of the decompressed bytes.
+	d := binary.PutUvarint(dst, uint64(len(src)))
+
+	// Return early if src is short.
+	if len(src) <= 4 {
+		if len(src) != 0 {
+			d += emitLiteral(dst[d:], src)
+		}
+		return dst[:d], nil
+	}
+
+	// Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive.
+	const maxTableSize = 1 << 14
+	shift, tableSize := uint(32-8), 1<<8
+	for tableSize < maxTableSize && tableSize < len(src) {
+		shift--
+		tableSize *= 2
+	}
+	var table [maxTableSize]int
+
+	// Iterate over the source bytes.
+	var (
+		s   int // The iterator position.
+		t   int // The last position with the same hash as s.
+		lit int // The start position of any pending literal bytes.
+	)
+	for s+3 < len(src) {
+		// Update the hash table.
+		b0, b1, b2, b3 := src[s], src[s+1], src[s+2], src[s+3]
+		h := uint32(b0) | uint32(b1)<<8 | uint32(b2)<<16 | uint32(b3)<<24
+		p := &table[(h*0x1e35a7bd)>>shift]
+		// We need to to store values in [-1, inf) in table. To save
+		// some initialization time, (re)use the table's zero value
+		// and shift the values against this zero: add 1 on writes,
+		// subtract 1 on reads.
+		t, *p = *p-1, s+1
+		// If t is invalid or src[s:s+4] differs from src[t:t+4], accumulate a literal byte.
+		if t < 0 || s-t >= maxOffset || b0 != src[t] || b1 != src[t+1] || b2 != src[t+2] || b3 != src[t+3] {
+			s++
+			continue
+		}
+		// Otherwise, we have a match. First, emit any pending literal bytes.
+		if lit != s {
+			d += emitLiteral(dst[d:], src[lit:s])
+		}
+		// Extend the match to be as long as possible.
+		s0 := s
+		s, t = s+4, t+4
+		for s < len(src) && src[s] == src[t] {
+			s++
+			t++
+		}
+		// Emit the copied bytes.
+		d += emitCopy(dst[d:], s-t, s-s0)
+		lit = s
+	}
+
+	// Emit any final pending literal bytes and return.
+	if lit != len(src) {
+		d += emitLiteral(dst[d:], src[lit:])
+	}
+	return dst[:d], nil
+}
+
+// MaxEncodedLen returns the maximum length of a snappy block, given its
+// uncompressed length.
+func MaxEncodedLen(srcLen int) int {
+	// Compressed data can be defined as:
+	//    compressed := item* literal*
+	//    item       := literal* copy
+	//
+	// The trailing literal sequence has a space blowup of at most 62/60
+	// since a literal of length 60 needs one tag byte + one extra byte
+	// for length information.
+	//
+	// Item blowup is trickier to measure. Suppose the "copy" op copies
+	// 4 bytes of data. Because of a special check in the encoding code,
+	// we produce a 4-byte copy only if the offset is < 65536. Therefore
+	// the copy op takes 3 bytes to encode, and this type of item leads
+	// to at most the 62/60 blowup for representing literals.
+	//
+	// Suppose the "copy" op copies 5 bytes of data. If the offset is big
+	// enough, it will take 5 bytes to encode the copy op. Therefore the
+	// worst case here is a one-byte literal followed by a five-byte copy.
+	// That is, 6 bytes of input turn into 7 bytes of "compressed" data.
+	//
+	// This last factor dominates the blowup, so the final estimate is:
+	return 32 + srcLen + srcLen/6
+}
+
+// NewWriter returns a new Writer that compresses to w, using the framing
+// format described at
+// https://code.google.com/p/snappy/source/browse/trunk/framing_format.txt
+func NewWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:   w,
+		enc: make([]byte, MaxEncodedLen(maxUncompressedChunkLen)),
+	}
+}
+
+// Writer is an io.Writer than can write Snappy-compressed bytes.
+type Writer struct {
+	w           io.Writer
+	err         error
+	enc         []byte
+	buf         [checksumSize + chunkHeaderSize]byte
+	wroteHeader bool
+}
+
+// Reset discards the writer's state and switches the Snappy writer to write to
+// w. This permits reusing a Writer rather than allocating a new one.
+func (w *Writer) Reset(writer io.Writer) {
+	w.w = writer
+	w.err = nil
+	w.wroteHeader = false
+}
+
+// Write satisfies the io.Writer interface.
+func (w *Writer) Write(p []byte) (n int, errRet error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+	if !w.wroteHeader {
+		copy(w.enc, magicChunk)
+		if _, err := w.w.Write(w.enc[:len(magicChunk)]); err != nil {
+			w.err = err
+			return n, err
+		}
+		w.wroteHeader = true
+	}
+	for len(p) > 0 {
+		var uncompressed []byte
+		if len(p) > maxUncompressedChunkLen {
+			uncompressed, p = p[:maxUncompressedChunkLen], p[maxUncompressedChunkLen:]
+		} else {
+			uncompressed, p = p, nil
+		}
+		checksum := crc(uncompressed)
+
+		// Compress the buffer, discarding the result if the improvement
+		// isn't at least 12.5%.
+		chunkType := uint8(chunkTypeCompressedData)
+		chunkBody, err := Encode(w.enc, uncompressed)
+		if err != nil {
+			w.err = err
+			return n, err
+		}
+		if len(chunkBody) >= len(uncompressed)-len(uncompressed)/8 {
+			chunkType, chunkBody = chunkTypeUncompressedData, uncompressed
+		}
+
+		chunkLen := 4 + len(chunkBody)
+		w.buf[0] = chunkType
+		w.buf[1] = uint8(chunkLen >> 0)
+		w.buf[2] = uint8(chunkLen >> 8)
+		w.buf[3] = uint8(chunkLen >> 16)
+		w.buf[4] = uint8(checksum >> 0)
+		w.buf[5] = uint8(checksum >> 8)
+		w.buf[6] = uint8(checksum >> 16)
+		w.buf[7] = uint8(checksum >> 24)
+		if _, err = w.w.Write(w.buf[:]); err != nil {
+			w.err = err
+			return n, err
+		}
+		if _, err = w.w.Write(chunkBody); err != nil {
+			w.err = err
+			return n, err
+		}
+		n += len(uncompressed)
+	}
+	return n, nil
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/snappy.go b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/snappy.go
new file mode 100644
index 000000000..043bf3d81
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/snappy.go
@@ -0,0 +1,68 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package snappy implements the snappy block-based compression format.
+// It aims for very high speeds and reasonable compression.
+//
+// The C++ snappy implementation is at http://code.google.com/p/snappy/
+package snappy
+
+import (
+	"hash/crc32"
+)
+
+/*
+Each encoded block begins with the varint-encoded length of the decoded data,
+followed by a sequence of chunks. Chunks begin and end on byte boundaries. The
+first byte of each chunk is broken into its 2 least and 6 most significant bits
+called l and m: l ranges in [0, 4) and m ranges in [0, 64). l is the chunk tag.
+Zero means a literal tag. All other values mean a copy tag.
+
+For literal tags:
+  - If m < 60, the next 1 + m bytes are literal bytes.
+  - Otherwise, let n be the little-endian unsigned integer denoted by the next
+    m - 59 bytes. The next 1 + n bytes after that are literal bytes.
+
+For copy tags, length bytes are copied from offset bytes ago, in the style of
+Lempel-Ziv compression algorithms. In particular:
+  - For l == 1, the offset ranges in [0, 1<<11) and the length in [4, 12).
+    The length is 4 + the low 3 bits of m. The high 3 bits of m form bits 8-10
+    of the offset. The next byte is bits 0-7 of the offset.
+  - For l == 2, the offset ranges in [0, 1<<16) and the length in [1, 65).
+    The length is 1 + m. The offset is the little-endian unsigned integer
+    denoted by the next 2 bytes.
+  - For l == 3, this tag is a legacy format that is no longer supported.
+*/
+const (
+	tagLiteral = 0x00
+	tagCopy1   = 0x01
+	tagCopy2   = 0x02
+	tagCopy4   = 0x03
+)
+
+const (
+	checksumSize    = 4
+	chunkHeaderSize = 4
+	magicChunk      = "\xff\x06\x00\x00" + magicBody
+	magicBody       = "sNaPpY"
+	// https://code.google.com/p/snappy/source/browse/trunk/framing_format.txt says
+	// that "the uncompressed data in a chunk must be no longer than 65536 bytes".
+	maxUncompressedChunkLen = 65536
+)
+
+const (
+	chunkTypeCompressedData   = 0x00
+	chunkTypeUncompressedData = 0x01
+	chunkTypePadding          = 0xfe
+	chunkTypeStreamIdentifier = 0xff
+)
+
+var crcTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc implements the checksum specified in section 3 of
+// https://code.google.com/p/snappy/source/browse/trunk/framing_format.txt
+func crc(b []byte) uint32 {
+	c := crc32.Update(0, crcTable, b)
+	return uint32(c>>15|c<<17) + 0xa282ead8
+}
diff --git a/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/snappy_test.go b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/snappy_test.go
new file mode 100644
index 000000000..0623385b7
--- /dev/null
+++ b/Godeps/_workspace/src/github.com/syndtr/gosnappy/snappy/snappy_test.go
@@ -0,0 +1,364 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"bytes"
+	"flag"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"math/rand"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+var (
+	download = flag.Bool("download", false, "If true, download any missing files before running benchmarks")
+	testdata = flag.String("testdata", "testdata", "Directory containing the test data")
+)
+
+func roundtrip(b, ebuf, dbuf []byte) error {
+	e, err := Encode(ebuf, b)
+	if err != nil {
+		return fmt.Errorf("encoding error: %v", err)
+	}
+	d, err := Decode(dbuf, e)
+	if err != nil {
+		return fmt.Errorf("decoding error: %v", err)
+	}
+	if !bytes.Equal(b, d) {
+		return fmt.Errorf("roundtrip mismatch:\n\twant %v\n\tgot  %v", b, d)
+	}
+	return nil
+}
+
+func TestEmpty(t *testing.T) {
+	if err := roundtrip(nil, nil, nil); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestSmallCopy(t *testing.T) {
+	for _, ebuf := range [][]byte{nil, make([]byte, 20), make([]byte, 64)} {
+		for _, dbuf := range [][]byte{nil, make([]byte, 20), make([]byte, 64)} {
+			for i := 0; i < 32; i++ {
+				s := "aaaa" + strings.Repeat("b", i) + "aaaabbbb"
+				if err := roundtrip([]byte(s), ebuf, dbuf); err != nil {
+					t.Errorf("len(ebuf)=%d, len(dbuf)=%d, i=%d: %v", len(ebuf), len(dbuf), i, err)
+				}
+			}
+		}
+	}
+}
+
+func TestSmallRand(t *testing.T) {
+	rng := rand.New(rand.NewSource(27354294))
+	for n := 1; n < 20000; n += 23 {
+		b := make([]byte, n)
+		for i := range b {
+			b[i] = uint8(rng.Uint32())
+		}
+		if err := roundtrip(b, nil, nil); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+func TestSmallRegular(t *testing.T) {
+	for n := 1; n < 20000; n += 23 {
+		b := make([]byte, n)
+		for i := range b {
+			b[i] = uint8(i%10 + 'a')
+		}
+		if err := roundtrip(b, nil, nil); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+func cmp(a, b []byte) error {
+	if len(a) != len(b) {
+		return fmt.Errorf("got %d bytes, want %d", len(a), len(b))
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return fmt.Errorf("byte #%d: got 0x%02x, want 0x%02x", i, a[i], b[i])
+		}
+	}
+	return nil
+}
+
+func TestFramingFormat(t *testing.T) {
+	// src is comprised of alternating 1e5-sized sequences of random
+	// (incompressible) bytes and repeated (compressible) bytes. 1e5 was chosen
+	// because it is larger than maxUncompressedChunkLen (64k).
+	src := make([]byte, 1e6)
+	rng := rand.New(rand.NewSource(1))
+	for i := 0; i < 10; i++ {
+		if i%2 == 0 {
+			for j := 0; j < 1e5; j++ {
+				src[1e5*i+j] = uint8(rng.Intn(256))
+			}
+		} else {
+			for j := 0; j < 1e5; j++ {
+				src[1e5*i+j] = uint8(i)
+			}
+		}
+	}
+
+	buf := new(bytes.Buffer)
+	if _, err := NewWriter(buf).Write(src); err != nil {
+		t.Fatalf("Write: encoding: %v", err)
+	}
+	dst, err := ioutil.ReadAll(NewReader(buf))
+	if err != nil {
+		t.Fatalf("ReadAll: decoding: %v", err)
+	}
+	if err := cmp(dst, src); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestReaderReset(t *testing.T) {
+	gold := bytes.Repeat([]byte("All that is gold does not glitter,\n"), 10000)
+	buf := new(bytes.Buffer)
+	if _, err := NewWriter(buf).Write(gold); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+	encoded, invalid, partial := buf.String(), "invalid", "partial"
+	r := NewReader(nil)
+	for i, s := range []string{encoded, invalid, partial, encoded, partial, invalid, encoded, encoded} {
+		if s == partial {
+			r.Reset(strings.NewReader(encoded))
+			if _, err := r.Read(make([]byte, 101)); err != nil {
+				t.Errorf("#%d: %v", i, err)
+				continue
+			}
+			continue
+		}
+		r.Reset(strings.NewReader(s))
+		got, err := ioutil.ReadAll(r)
+		switch s {
+		case encoded:
+			if err != nil {
+				t.Errorf("#%d: %v", i, err)
+				continue
+			}
+			if err := cmp(got, gold); err != nil {
+				t.Errorf("#%d: %v", i, err)
+				continue
+			}
+		case invalid:
+			if err == nil {
+				t.Errorf("#%d: got nil error, want non-nil", i)
+				continue
+			}
+		}
+	}
+}
+
+func TestWriterReset(t *testing.T) {
+	gold := bytes.Repeat([]byte("Not all those who wander are lost;\n"), 10000)
+	var gots, wants [][]byte
+	const n = 20
+	w, failed := NewWriter(nil), false
+	for i := 0; i <= n; i++ {
+		buf := new(bytes.Buffer)
+		w.Reset(buf)
+		want := gold[:len(gold)*i/n]
+		if _, err := w.Write(want); err != nil {
+			t.Errorf("#%d: Write: %v", i, err)
+			failed = true
+			continue
+		}
+		got, err := ioutil.ReadAll(NewReader(buf))
+		if err != nil {
+			t.Errorf("#%d: ReadAll: %v", i, err)
+			failed = true
+			continue
+		}
+		gots = append(gots, got)
+		wants = append(wants, want)
+	}
+	if failed {
+		return
+	}
+	for i := range gots {
+		if err := cmp(gots[i], wants[i]); err != nil {
+			t.Errorf("#%d: %v", i, err)
+		}
+	}
+}
+
+func benchDecode(b *testing.B, src []byte) {
+	encoded, err := Encode(nil, src)
+	if err != nil {
+		b.Fatal(err)
+	}
+	// Bandwidth is in amount of uncompressed data.
+	b.SetBytes(int64(len(src)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		Decode(src, encoded)
+	}
+}
+
+func benchEncode(b *testing.B, src []byte) {
+	// Bandwidth is in amount of uncompressed data.
+	b.SetBytes(int64(len(src)))
+	dst := make([]byte, MaxEncodedLen(len(src)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		Encode(dst, src)
+	}
+}
+
+func readFile(b testing.TB, filename string) []byte {
+	src, err := ioutil.ReadFile(filename)
+	if err != nil {
+		b.Fatalf("failed reading %s: %s", filename, err)
+	}
+	if len(src) == 0 {
+		b.Fatalf("%s has zero length", filename)
+	}
+	return src
+}
+
+// expand returns a slice of length n containing repeated copies of src.
+func expand(src []byte, n int) []byte {
+	dst := make([]byte, n)
+	for x := dst; len(x) > 0; {
+		i := copy(x, src)
+		x = x[i:]
+	}
+	return dst
+}
+
+func benchWords(b *testing.B, n int, decode bool) {
+	// Note: the file is OS-language dependent so the resulting values are not
+	// directly comparable for non-US-English OS installations.
+	data := expand(readFile(b, "/usr/share/dict/words"), n)
+	if decode {
+		benchDecode(b, data)
+	} else {
+		benchEncode(b, data)
+	}
+}
+
+func BenchmarkWordsDecode1e3(b *testing.B) { benchWords(b, 1e3, true) }
+func BenchmarkWordsDecode1e4(b *testing.B) { benchWords(b, 1e4, true) }
+func BenchmarkWordsDecode1e5(b *testing.B) { benchWords(b, 1e5, true) }
+func BenchmarkWordsDecode1e6(b *testing.B) { benchWords(b, 1e6, true) }
+func BenchmarkWordsEncode1e3(b *testing.B) { benchWords(b, 1e3, false) }
+func BenchmarkWordsEncode1e4(b *testing.B) { benchWords(b, 1e4, false) }
+func BenchmarkWordsEncode1e5(b *testing.B) { benchWords(b, 1e5, false) }
+func BenchmarkWordsEncode1e6(b *testing.B) { benchWords(b, 1e6, false) }
+
+// testFiles' values are copied directly from
+// https://raw.githubusercontent.com/google/snappy/master/snappy_unittest.cc
+// The label field is unused in snappy-go.
+var testFiles = []struct {
+	label    string
+	filename string
+}{
+	{"html", "html"},
+	{"urls", "urls.10K"},
+	{"jpg", "fireworks.jpeg"},
+	{"jpg_200", "fireworks.jpeg"},
+	{"pdf", "paper-100k.pdf"},
+	{"html4", "html_x_4"},
+	{"txt1", "alice29.txt"},
+	{"txt2", "asyoulik.txt"},
+	{"txt3", "lcet10.txt"},
+	{"txt4", "plrabn12.txt"},
+	{"pb", "geo.protodata"},
+	{"gaviota", "kppkn.gtb"},
+}
+
+// The test data files are present at this canonical URL.
+const baseURL = "https://raw.githubusercontent.com/google/snappy/master/testdata/"
+
+func downloadTestdata(basename string) (errRet error) {
+	filename := filepath.Join(*testdata, basename)
+	if stat, err := os.Stat(filename); err == nil && stat.Size() != 0 {
+		return nil
+	}
+
+	if !*download {
+		return fmt.Errorf("test data not found; skipping benchmark without the -download flag")
+	}
+	// Download the official snappy C++ implementation reference test data
+	// files for benchmarking.
+	if err := os.Mkdir(*testdata, 0777); err != nil && !os.IsExist(err) {
+		return fmt.Errorf("failed to create testdata: %s", err)
+	}
+
+	f, err := os.Create(filename)
+	if err != nil {
+		return fmt.Errorf("failed to create %s: %s", filename, err)
+	}
+	defer f.Close()
+	defer func() {
+		if errRet != nil {
+			os.Remove(filename)
+		}
+	}()
+	url := baseURL + basename
+	resp, err := http.Get(url)
+	if err != nil {
+		return fmt.Errorf("failed to download %s: %s", url, err)
+	}
+	defer resp.Body.Close()
+	if s := resp.StatusCode; s != http.StatusOK {
+		return fmt.Errorf("downloading %s: HTTP status code %d (%s)", url, s, http.StatusText(s))
+	}
+	_, err = io.Copy(f, resp.Body)
+	if err != nil {
+		return fmt.Errorf("failed to download %s to %s: %s", url, filename, err)
+	}
+	return nil
+}
+
+func benchFile(b *testing.B, n int, decode bool) {
+	if err := downloadTestdata(testFiles[n].filename); err != nil {
+		b.Fatalf("failed to download testdata: %s", err)
+	}
+	data := readFile(b, filepath.Join(*testdata, testFiles[n].filename))
+	if decode {
+		benchDecode(b, data)
+	} else {
+		benchEncode(b, data)
+	}
+}
+
+// Naming convention is kept similar to what snappy's C++ implementation uses.
+func Benchmark_UFlat0(b *testing.B)  { benchFile(b, 0, true) }
+func Benchmark_UFlat1(b *testing.B)  { benchFile(b, 1, true) }
+func Benchmark_UFlat2(b *testing.B)  { benchFile(b, 2, true) }
+func Benchmark_UFlat3(b *testing.B)  { benchFile(b, 3, true) }
+func Benchmark_UFlat4(b *testing.B)  { benchFile(b, 4, true) }
+func Benchmark_UFlat5(b *testing.B)  { benchFile(b, 5, true) }
+func Benchmark_UFlat6(b *testing.B)  { benchFile(b, 6, true) }
+func Benchmark_UFlat7(b *testing.B)  { benchFile(b, 7, true) }
+func Benchmark_UFlat8(b *testing.B)  { benchFile(b, 8, true) }
+func Benchmark_UFlat9(b *testing.B)  { benchFile(b, 9, true) }
+func Benchmark_UFlat10(b *testing.B) { benchFile(b, 10, true) }
+func Benchmark_UFlat11(b *testing.B) { benchFile(b, 11, true) }
+func Benchmark_ZFlat0(b *testing.B)  { benchFile(b, 0, false) }
+func Benchmark_ZFlat1(b *testing.B)  { benchFile(b, 1, false) }
+func Benchmark_ZFlat2(b *testing.B)  { benchFile(b, 2, false) }
+func Benchmark_ZFlat3(b *testing.B)  { benchFile(b, 3, false) }
+func Benchmark_ZFlat4(b *testing.B)  { benchFile(b, 4, false) }
+func Benchmark_ZFlat5(b *testing.B)  { benchFile(b, 5, false) }
+func Benchmark_ZFlat6(b *testing.B)  { benchFile(b, 6, false) }
+func Benchmark_ZFlat7(b *testing.B)  { benchFile(b, 7, false) }
+func Benchmark_ZFlat8(b *testing.B)  { benchFile(b, 8, false) }
+func Benchmark_ZFlat9(b *testing.B)  { benchFile(b, 9, false) }
+func Benchmark_ZFlat10(b *testing.B) { benchFile(b, 10, false) }
+func Benchmark_ZFlat11(b *testing.B) { benchFile(b, 11, false) }