Improved read/write performance by another ~2x by not calling Stat() on every read/write

README.md

@@ -97,31 +97,33 @@ Benchmarks run on a 11" Macbook with a 1.4Ghz Intel Core i7:
 ```
 $ make bench
 ...
-BenchmarkGet/128B-4         	  200000	      5780 ns/op	     400 B/op	       5 allocs/op
+BenchmarkGet/128B-4         	  300000	      5178 ns/op	     400 B/op	       5 allocs/op
-BenchmarkGet/256B-4         	  200000	      6138 ns/op	     656 B/op	       5 allocs/op
+BenchmarkGet/256B-4         	  300000	      5273 ns/op	     656 B/op	       5 allocs/op
-BenchmarkGet/512B-4         	  200000	      5967 ns/op	    1200 B/op	       5 allocs/op
+BenchmarkGet/512B-4         	  200000	      5368 ns/op	    1200 B/op	       5 allocs/op
-BenchmarkGet/1K-4           	  200000	      6290 ns/op	    2288 B/op	       5 allocs/op
+BenchmarkGet/1K-4           	  200000	      5800 ns/op	    2288 B/op	       5 allocs/op
-BenchmarkGet/2K-4           	  200000	      6293 ns/op	    4464 B/op	       5 allocs/op
+BenchmarkGet/2K-4           	  200000	      6766 ns/op	    4464 B/op	       5 allocs/op
-BenchmarkGet/4K-4           	  200000	      7673 ns/op	    9072 B/op	       5 allocs/op
+BenchmarkGet/4K-4           	  200000	      7857 ns/op	    9072 B/op	       5 allocs/op
-BenchmarkGet/8K-4           	  200000	     10373 ns/op	   17776 B/op	       5 allocs/op
+BenchmarkGet/8K-4           	  200000	      9538 ns/op	   17776 B/op	       5 allocs/op
-BenchmarkGet/16K-4          	  100000	     14227 ns/op	   34928 B/op	       5 allocs/op
+BenchmarkGet/16K-4          	  100000	     13188 ns/op	   34928 B/op	       5 allocs/op
-BenchmarkGet/32K-4          	  100000	     25953 ns/op	   73840 B/op	       5 allocs/op
+BenchmarkGet/32K-4          	  100000	     21620 ns/op	   73840 B/op	       5 allocs/op
 
-BenchmarkPut/128B-4         	  100000	     17353 ns/op	     680 B/op	       5 allocs/op
+BenchmarkPut/128B-4         	  200000	      7875 ns/op	     409 B/op	       6 allocs/op
-BenchmarkPut/256B-4         	  100000	     18620 ns/op	     808 B/op	       5 allocs/op
+BenchmarkPut/256B-4         	  200000	      8712 ns/op	     538 B/op	       6 allocs/op
-BenchmarkPut/512B-4         	  100000	     19068 ns/op	    1096 B/op	       5 allocs/op
+BenchmarkPut/512B-4         	  200000	      9832 ns/op	     829 B/op	       6 allocs/op
-BenchmarkPut/1K-4           	  100000	     23738 ns/op	    1673 B/op	       5 allocs/op
+BenchmarkPut/1K-4           	  100000	     13105 ns/op	    1410 B/op	       6 allocs/op
-BenchmarkPut/2K-4           	   50000	     25118 ns/op	    2826 B/op	       5 allocs/op
+BenchmarkPut/2K-4           	  100000	     18601 ns/op	    2572 B/op	       6 allocs/op
-BenchmarkPut/4K-4           	   50000	     44605 ns/op	    5389 B/op	       5 allocs/op
+BenchmarkPut/4K-4           	   50000	     36631 ns/op	    5151 B/op	       6 allocs/op
-BenchmarkPut/8K-4           	   30000	     55237 ns/op	   10001 B/op	       5 allocs/op
+BenchmarkPut/8K-4           	   30000	     56128 ns/op	    9798 B/op	       6 allocs/op
-BenchmarkPut/16K-4          	   20000	     78966 ns/op	   18972 B/op	       5 allocs/op
+BenchmarkPut/16K-4          	   20000	     83209 ns/op	   18834 B/op	       6 allocs/op
-BenchmarkPut/32K-4          	   10000	    116253 ns/op	   41520 B/op	       5 allocs/op
+BenchmarkPut/32K-4          	   10000	    135899 ns/op	   41517 B/op	       6 allocs/op
+
+BenchmarkScan-4             	 1000000	      1851 ns/op	     493 B/op	      25 allocs/op
 ```
 
 For 128B values:
 
-* ~180,000 reads/sec
+* ~200,000 reads/sec
-*  ~60,000 writes/sec
+* ~130,000 writes/sec
 
 The full benchmark above shows linear performance as you increase key/value sizes.
 

bitcask.go

@@ -9,6 +9,7 @@ import (
 	"strings"
 	"time"
 
+	"github.com/derekparker/trie"
 	"github.com/gofrs/flock"
 )
 
@@ -27,6 +28,7 @@ type Bitcask struct {
 	curr      *Datafile
 	keydir    *Keydir
 	datafiles []*Datafile
+	trie      *trie.Trie
 
 	maxDatafileSize int64
 }
@@ -82,7 +84,8 @@ func (b *Bitcask) Put(key string, value []byte) error {
 		return err
 	}
 
-	b.keydir.Add(key, b.curr.id, index, time.Now().Unix())
+	item := b.keydir.Add(key, b.curr.id, index, time.Now().Unix())
+	b.trie.Add(key, item)
 
 	return nil
 }
@@ -94,10 +97,21 @@ func (b *Bitcask) Delete(key string) error {
 	}
 
 	b.keydir.Delete(key)
+	b.trie.Remove(key)
 
 	return nil
 }
 
+func (b *Bitcask) Scan(prefix string, f func(key string) error) error {
+	keys := b.trie.PrefixSearch(prefix)
+	for _, key := range keys {
+		if err := f(key); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 func (b *Bitcask) Fold(f func(key string) error) error {
 	for key := range b.keydir.Keys() {
 		if err := f(key); err != nil {
@@ -265,9 +279,11 @@ func Open(path string, options ...func(*Bitcask) error) (*Bitcask, error) {
 		return nil, err
 	}
 
-	keydir := NewKeydir()
 	var datafiles []*Datafile
 
+	keydir := NewKeydir()
+	trie := trie.New()
+
 	for i, fn := range fns {
 		df, err := NewDatafile(path, ids[i], true)
 		if err != nil {
@@ -289,7 +305,8 @@ func Open(path string, options ...func(*Bitcask) error) (*Bitcask, error) {
 
 			for key := range hint.Keys() {
 				item, _ := hint.Get(key)
-				keydir.Add(key, item.FileID, item.Index, item.Timestamp)
+				_ = keydir.Add(key, item.FileID, item.Index, item.Timestamp)
+				trie.Add(key, item)
 			}
 		} else {
 			for {
@@ -307,7 +324,8 @@ func Open(path string, options ...func(*Bitcask) error) (*Bitcask, error) {
 					continue
 				}
 
-				keydir.Add(e.Key, ids[i], e.Index, e.Timestamp)
+				item := keydir.Add(e.Key, ids[i], e.Index, e.Timestamp)
+				trie.Add(e.Key, item)
 			}
 		}
 	}
@@ -329,6 +347,7 @@ func Open(path string, options ...func(*Bitcask) error) (*Bitcask, error) {
 		curr:      curr,
 		keydir:    keydir,
 		datafiles: datafiles,
+		trie:      trie,
 
 		maxDatafileSize: DefaultMaxDatafileSize,
 	}

bitcask_test.go

@@ -3,6 +3,8 @@ package bitcask
 import (
 	"fmt"
 	"io/ioutil"
+	"reflect"
+	"sort"
 	"strings"
 	"sync"
 	"testing"
@@ -327,6 +329,58 @@ func TestConcurrent(t *testing.T) {
 	})
 }
 
+func TestScan(t *testing.T) {
+	assert := assert.New(t)
+
+	testdir, err := ioutil.TempDir("", "bitcask")
+	assert.NoError(err)
+
+	var db *Bitcask
+
+	t.Run("Setup", func(t *testing.T) {
+		t.Run("Open", func(t *testing.T) {
+			db, err = Open(testdir)
+			assert.NoError(err)
+		})
+
+		t.Run("Put", func(t *testing.T) {
+			var items = map[string][]byte{
+				"1":     []byte("1"),
+				"2":     []byte("2"),
+				"3":     []byte("3"),
+				"food":  []byte("pizza"),
+				"foo":   []byte("foo"),
+				"fooz":  []byte("fooz ball"),
+				"hello": []byte("world"),
+			}
+			for k, v := range items {
+				err = db.Put(k, v)
+				assert.NoError(err)
+			}
+		})
+	})
+
+	t.Run("Scan", func(t *testing.T) {
+		var (
+			vals     []string
+			expected = []string{
+				"foo",
+				"fooz ball",
+				"pizza",
+			}
+		)
+
+		err = db.Scan("fo", func(key string) error {
+			val, err := db.Get(key)
+			assert.NoError(err)
+			vals = append(vals, string(val))
+			return nil
+		})
+		sort.Strings(vals)
+		assert.Equal(expected, vals)
+	})
+}
+
 func TestLocking(t *testing.T) {
 	assert := assert.New(t)
 
@@ -433,3 +487,47 @@ func BenchmarkPut(b *testing.B) {
 		})
 	}
 }
+
+func BenchmarkScan(b *testing.B) {
+	testdir, err := ioutil.TempDir("", "bitcask")
+	if err != nil {
+		b.Fatal(err)
+	}
+
+	db, err := Open(testdir)
+	if err != nil {
+		b.Fatal(err)
+	}
+	defer db.Close()
+
+	var items = map[string][]byte{
+		"1":     []byte("1"),
+		"2":     []byte("2"),
+		"3":     []byte("3"),
+		"food":  []byte("pizza"),
+		"foo":   []byte("foo"),
+		"fooz":  []byte("fooz ball"),
+		"hello": []byte("world"),
+	}
+	for k, v := range items {
+		err := db.Put(k, v)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+
+	var expected = []string{"foo", "food", "fooz"}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		var keys []string
+		err = db.Scan("fo", func(key string) error {
+			keys = append(keys, key)
+			return nil
+		})
+		sort.Strings(keys)
+		if !reflect.DeepEqual(expected, keys) {
+			b.Fatal(fmt.Errorf("expected keys=#%v got=%#v", expected, keys))
+		}
+	}
+}

cmd/bitcask/scan.go

@@ -0,0 +1,60 @@
+package main
+
+import (
+	"fmt"
+	"os"
+
+	log "github.com/sirupsen/logrus"
+	"github.com/spf13/cobra"
+	"github.com/spf13/viper"
+
+	"github.com/prologic/bitcask"
+)
+
+var scanCmd = &cobra.Command{
+	Use:     "scan <prefix>",
+	Aliases: []string{"search", "find"},
+	Short:   "Perform a prefis scan for keys",
+	Long: `This performa a prefix scan for keys  starting with the given
+prefix. This uses a Trie to search for matching keys and returns all matched
+keys.`,
+	Args: cobra.ExactArgs(1),
+	Run: func(cmd *cobra.Command, args []string) {
+		path := viper.GetString("path")
+
+		prefix := args[0]
+
+		os.Exit(scan(path, prefix))
+	},
+}
+
+func init() {
+	RootCmd.AddCommand(scanCmd)
+}
+
+func scan(path, prefix string) int {
+	db, err := bitcask.Open(path)
+	if err != nil {
+		log.WithError(err).Error("error opening database")
+		return 1
+	}
+	defer db.Close()
+
+	err = db.Scan(prefix, func(key string) error {
+		value, err := db.Get(key)
+		if err != nil {
+			log.WithError(err).Error("error reading key")
+			return err
+		}
+
+		fmt.Printf("%s\n", string(value))
+		log.WithField("key", key).WithField("value", value).Debug("key/value")
+		return nil
+	})
+	if err != nil {
+		log.WithError(err).Error("error scanning keys")
+		return 1
+	}
+
+	return 0
+}

datafile.go

@@ -1,13 +1,14 @@
 package bitcask
 
 import (
-	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
 	"sync"
 	"time"
 
+	"github.com/pkg/errors"
+
 	pb "github.com/prologic/bitcask/proto"
 	"github.com/prologic/bitcask/streampb"
 )
@@ -23,11 +24,12 @@ var (
 type Datafile struct {
 	sync.RWMutex
 
-	id  int
+	id     int
-	r   *os.File
+	r      *os.File
-	w   *os.File
+	w      *os.File
-	dec *streampb.Decoder
+	offset int64
-	enc *streampb.Encoder
+	dec    *streampb.Decoder
+	enc    *streampb.Encoder
 }
 
 func NewDatafile(path string, id int, readonly bool) (*Datafile, error) {
@@ -50,16 +52,23 @@ func NewDatafile(path string, id int, readonly bool) (*Datafile, error) {
 	if err != nil {
 		return nil, err
 	}
+	stat, err := r.Stat()
+	if err != nil {
+		return nil, errors.Wrap(err, "error calling Stat()")
+	}
+
+	offset := stat.Size()
 
 	dec := streampb.NewDecoder(r)
 	enc := streampb.NewEncoder(w)
 
 	return &Datafile{
-		id:  id,
+		id:     id,
-		r:   r,
+		r:      r,
-		w:   w,
+		w:      w,
-		dec: dec,
+		offset: offset,
-		enc: enc,
+		dec:    dec,
+		enc:    enc,
 	}, nil
 }
 
@@ -87,22 +96,7 @@ func (df *Datafile) Sync() error {
 }
 
 func (df *Datafile) Size() (int64, error) {
-	var (
+	return df.offset, nil
-		stat os.FileInfo
-		err  error
-	)
-
-	if df.w == nil {
-		stat, err = df.r.Stat()
-	} else {
-		stat, err = df.w.Stat()
-	}
-
-	if err != nil {
-		return -1, err
-	}
-
-	return stat.Size(), nil
 }
 
 func (df *Datafile) Read() (e pb.Entry, err error) {
@@ -129,23 +123,17 @@ func (df *Datafile) Write(e pb.Entry) (int64, error) {
 		return -1, ErrReadonly
 	}
 
-	stat, err := df.w.Stat()
+	e.Index = df.offset
-	if err != nil {
-		return -1, err
-	}
-
-	index := stat.Size()
-
-	e.Index = index
 	e.Timestamp = time.Now().Unix()
 
 	df.Lock()
-	err = df.enc.Encode(&e)
+	n, err := df.enc.Encode(&e)
 	df.Unlock()
 
 	if err != nil {
 		return -1, err
 	}
+	df.offset += n
 
-	return index, nil
+	return e.Index, nil
 }

go.mod

@@ -1,6 +1,7 @@
 module github.com/prologic/bitcask
 
 require (
+	github.com/derekparker/trie v0.0.0-20180212171413-e608c2733dc7
 	github.com/gofrs/flock v0.7.1
 	github.com/gogo/protobuf v1.2.1
 	github.com/golang/protobuf v1.2.0

go.sum

@@ -8,6 +8,8 @@ github.com/davecgh/go-spew v1.1.0 h1:ZDRjVQ15GmhC3fiQ8ni8+OwkZQO4DARzQgrnXU1Liz8
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/derekparker/trie v0.0.0-20180212171413-e608c2733dc7 h1:Cab9yoTQh1TxObKfis1DzZ6vFLK5kbeenMjRES/UE3o=
+github.com/derekparker/trie v0.0.0-20180212171413-e608c2733dc7/go.mod h1:D6ICZm05D9VN1n/8iOtBxLpXtoGp6HDFUJ1RNVieOSE=
 github.com/fsnotify/fsnotify v1.4.7 h1:IXs+QLmnXW2CcXuY+8Mzv/fWEsPGWxqefPtCP5CnV9I=
 github.com/fsnotify/fsnotify v1.4.7/go.mod h1:jwhsz4b93w/PPRr/qN1Yymfu8t87LnFCMoQvtojpjFo=
 github.com/gofrs/flock v0.7.1 h1:DP+LD/t0njgoPBvT5MJLeliUIVQR03hiKR6vezdwHlc=

keydir.go

@@ -25,15 +25,18 @@ func NewKeydir() *Keydir {
 	}
 }
 
-func (k *Keydir) Add(key string, fileid int, index, timestamp int64) {
+func (k *Keydir) Add(key string, fileid int, index, timestamp int64) Item {
-	k.Lock()
+	item := Item{
-	defer k.Unlock()
-
-	k.kv[key] = Item{
 		FileID:    fileid,
 		Index:     index,
 		Timestamp: timestamp,
 	}
+
+	k.Lock()
+	k.kv[key] = item
+	k.Unlock()
+
+	return item
 }
 
 func (k *Keydir) Get(key string) (Item, bool) {

streampb/stream.go

@@ -1,6 +1,7 @@
 package streampb
 
 import (
+	"bufio"
 	"encoding/binary"
 	"io"
 
@@ -16,32 +17,40 @@ const (
 
 // NewEncoder creates a streaming protobuf encoder.
 func NewEncoder(w io.Writer) *Encoder {
-	return &Encoder{w}
+	return &Encoder{w: bufio.NewWriter(w)}
 }
 
 // Encoder wraps an underlying io.Writer and allows you to stream
 // proto encodings on it.
 type Encoder struct {
-	w io.Writer
+	w *bufio.Writer
 }
 
 // Encode takes any proto.Message and streams it to the underlying writer.
 // Messages are framed with a length prefix.
-func (e *Encoder) Encode(msg proto.Message) error {
+func (e *Encoder) Encode(msg proto.Message) (int64, error) {
 	prefixBuf := make([]byte, prefixSize)
 
 	buf, err := proto.Marshal(msg)
 	if err != nil {
-		return err
+		return 0, err
 	}
 	binary.BigEndian.PutUint64(prefixBuf, uint64(len(buf)))
 
 	if _, err := e.w.Write(prefixBuf); err != nil {
-		return errors.Wrap(err, "failed writing length prefix")
+		return 0, errors.Wrap(err, "failed writing length prefix")
 	}
 
-	_, err = e.w.Write(buf)
+	n, err := e.w.Write(buf)
-	return errors.Wrap(err, "failed writing marshaled data")
+	if err != nil {
+		return 0, errors.Wrap(err, "failed writing marshaled data")
+	}
+
+	if err = e.w.Flush(); err != nil {
+		return 0, errors.Wrap(err, "failed flushing data")
+	}
+
+	return int64(n + prefixSize), nil
 }
 
 // NewDecoder creates a streaming protobuf decoder.