// makeRoomForWrite ensures that there is room in d.mem for the next write.
//
// d.mu must be held when calling this, but the mutex may be dropped and
// re-acquired during the course of this method.
func (d *DB) makeRoomForWrite(force bool) error {
allowDelay := !force
for {
// TODO: check any previous sticky error, if the paranoid option is set.
if allowDelay && len(d.versions.currentVersion().files[0]) > l0SlowdownWritesTrigger {
// We are getting close to hitting a hard limit on the number of
// L0 files. Rather than delaying a single write by several
// seconds when we hit the hard limit, start delaying each
// individual write by 1ms to reduce latency variance.
d.mu.Unlock()
time.Sleep(1 * time.Millisecond)
d.mu.Lock()
allowDelay = false
// TODO: how do we ensure we are still 'at the front of the writer queue'?
continue
}
if !force && d.mem.ApproximateMemoryUsage() <= d.opts.GetWriteBufferSize() {
// There is room in the current memtable.
break
}
if d.imm != nil {
// We have filled up the current memtable, but the previous
// one is still being compacted, so we wait.
d.compactionCond.Wait()
continue
}
if len(d.versions.currentVersion().files[0]) > l0StopWritesTrigger {
// There are too many level-0 files.
d.compactionCond.Wait()
continue
}
// Attempt to switch to a new memtable and trigger compaction of old
// TODO: drop and re-acquire d.mu around the I/O.
newLogNumber := d.versions.nextFileNum()
newLogFile, err := d.opts.GetFileSystem().Create(dbFilename(d.dirname, fileTypeLog, newLogNumber))
if err != nil {
return err
}
newLog := record.NewWriter(newLogFile)
if err := d.log.Close(); err != nil {
newLogFile.Close()
return err
}
if err := d.logFile.Close(); err != nil {
newLog.Close()
newLogFile.Close()
return err
}
d.logNumber, d.logFile, d.log = newLogNumber, newLogFile, newLog
d.imm, d.mem = d.mem, memdb.New(&d.icmpOpts)
force = false
d.maybeScheduleCompaction()
}
return nil
}
// Open opens a LevelDB whose files live in the given directory.
func Open(dirname string, opts *db.Options) (*DB, error) {
d := &DB{
dirname: dirname,
opts: opts,
icmp: internalKeyComparer{opts.GetComparer()},
pendingOutputs: make(map[uint64]struct{}),
}
if opts != nil {
d.icmpOpts = *opts
}
d.icmpOpts.Comparer = d.icmp
tableCacheSize := opts.GetMaxOpenFiles() - numNonTableCacheFiles
if tableCacheSize < minTableCacheSize {
tableCacheSize = minTableCacheSize
}
d.tableCache.init(dirname, opts.GetFileSystem(), &d.icmpOpts, tableCacheSize)
d.mem = memdb.New(&d.icmpOpts)
d.compactionCond = sync.Cond{L: &d.mu}
fs := opts.GetFileSystem()
d.mu.Lock()
defer d.mu.Unlock()
// Lock the database directory.
err := fs.MkdirAll(dirname, 0755)
if err != nil {
return nil, err
}
fileLock, err := fs.Lock(dbFilename(dirname, fileTypeLock, 0))
if err != nil {
return nil, err
}
defer func() {
if fileLock != nil {
fileLock.Close()
}
}()
if _, err := fs.Stat(dbFilename(dirname, fileTypeCurrent, 0)); os.IsNotExist(err) {
// Create the DB if it did not already exist.
if err := createDB(dirname, opts); err != nil {
return nil, err
}
} else if err != nil {
return nil, fmt.Errorf("leveldb: database %q: %v", dirname, err)
} else if opts.GetErrorIfDBExists() {
return nil, fmt.Errorf("leveldb: database %q already exists", dirname)
}
// Load the version set.
err = d.versions.load(dirname, opts)
if err != nil {
return nil, err
}
// Replay any newer log files than the ones named in the manifest.
var ve versionEdit
ls, err := fs.List(dirname)
if err != nil {
return nil, err
}
var logFiles fileNumAndNameSlice
for _, filename := range ls {
ft, fn, ok := parseDBFilename(filename)
if ok && ft == fileTypeLog && (fn >= d.versions.logNumber || fn == d.versions.prevLogNumber) {
logFiles = append(logFiles, fileNumAndName{fn, filename})
}
}
sort.Sort(logFiles)
for _, lf := range logFiles {
maxSeqNum, err := d.replayLogFile(&ve, fs, filepath.Join(dirname, lf.name))
if err != nil {
return nil, err
}
d.versions.markFileNumUsed(lf.num)
if d.versions.lastSequence < maxSeqNum {
d.versions.lastSequence = maxSeqNum
}
}
// Create an empty .log file.
ve.logNumber = d.versions.nextFileNum()
d.logNumber = ve.logNumber
logFile, err := fs.Create(dbFilename(dirname, fileTypeLog, ve.logNumber))
if err != nil {
return nil, err
}
defer func() {
if logFile != nil {
logFile.Close()
}
}()
d.log = record.NewWriter(logFile)
// Write a new manifest to disk.
if err := d.versions.logAndApply(dirname, &ve); err != nil {
return nil, err
}
d.deleteObsoleteFiles()
d.maybeScheduleCompaction()
d.logFile, logFile = logFile, nil
d.fileLock, fileLock = fileLock, nil
return d, nil
}
// replayLogFile replays the edits in the named log file.
//
// d.mu must be held when calling this, but the mutex may be dropped and
// re-acquired during the course of this method.
func (d *DB) replayLogFile(ve *versionEdit, fs db.FileSystem, filename string) (maxSeqNum uint64, err error) {
file, err := fs.Open(filename)
if err != nil {
return 0, err
}
defer file.Close()
var (
mem *memdb.MemDB
batchBuf = new(bytes.Buffer)
ikey = make(internalKey, 512)
rr = record.NewReader(file)
)
for {
r, err := rr.Next()
if err == io.EOF {
break
}
if err != nil {
return 0, err
}
_, err = io.Copy(batchBuf, r)
if err != nil {
return 0, err
}
if batchBuf.Len() < batchHeaderLen {
return 0, fmt.Errorf("leveldb: corrupt log file %q", filename)
}
b := Batch{batchBuf.Bytes()}
seqNum := b.seqNum()
seqNum1 := seqNum + uint64(b.count())
if maxSeqNum < seqNum1 {
maxSeqNum = seqNum1
}
if mem == nil {
mem = memdb.New(&d.icmpOpts)
}
t := b.iter()
for ; seqNum != seqNum1; seqNum++ {
kind, ukey, value, ok := t.next()
if !ok {
return 0, fmt.Errorf("leveldb: corrupt log file %q", filename)
}
// Convert seqNum, kind and key into an internalKey, and add that ikey/value
// pair to mem.
//
// TODO: instead of copying to an intermediate buffer (ikey), is it worth
// adding a SetTwoPartKey(db.TwoPartKey{key0, key1}, value, opts) method to
// memdb.MemDB? What effect does that have on the db.Comparer interface?
//
// The C++ LevelDB code does not need an intermediate copy because its memdb
// implementation is a private implementation detail, and copies each internal
// key component from the Batch format straight to the skiplist buffer.
//
// Go's LevelDB considers the memdb functionality to be useful in its own
// right, and so leveldb/memdb is a separate package that is usable without
// having to import the top-level leveldb package. That extra abstraction
// means that we need to copy to an intermediate buffer here, to reconstruct
// the complete internal key to pass to the memdb.
ikey = makeInternalKey(ikey, ukey, kind, seqNum)
mem.Set(ikey, value, nil)
}
if len(t) != 0 {
return 0, fmt.Errorf("leveldb: corrupt log file %q", filename)
}
// TODO: if mem is large enough, write it to a level-0 table and set mem = nil.
batchBuf.Reset()
}
if mem != nil && !mem.Empty() {
meta, err := d.writeLevel0Table(fs, mem)
if err != nil {
return 0, err
}
ve.newFiles = append(ve.newFiles, newFileEntry{level: 0, meta: meta})
// Strictly speaking, it's too early to delete meta.fileNum from d.pendingOutputs,
// but we are replaying the log file, which happens before Open returns, so there
// is no possibility of deleteObsoleteFiles being called concurrently here.
delete(d.pendingOutputs, meta.fileNum)
}
return maxSeqNum, nil
}
func TestVersion(t *testing.T) {
// testTable is a table to insert into a version.
// Each element of data is a string of the form "internalKey value".
type testTable struct {
level int
fileNum uint64
data []string
}
testCases := []struct {
description string
// badOrdering is whether this test case has a table ordering violation.
badOrdering bool
// tables are the tables to populate the version with.
tables []testTable
// queries are the queries to run against the version. Each element has
// the form "internalKey wantedValue". The internalKey is passed to the
// version.get method, wantedValue may be "ErrNotFound" if the query
// should return that error.
queries []string
}{
{
description: "empty: an empty version",
queries: []string{
"abc.MAX.101 ErrNotFound",
},
},
{
description: "single-0: one level-0 table",
tables: []testTable{
{
level: 0,
fileNum: 10,
data: []string{
"the.SET.101 a",
"cat.SET.102 b",
"on_.SET.103 c",
"the.SET.104 d",
"mat.SET.105 e",
"the.DEL.106 ",
"the.SET.107 g",
},
},
},
queries: []string{
"aaa.MAX.105 ErrNotFound",
"cat.MAX.105 b",
"hat.MAX.105 ErrNotFound",
"mat.MAX.105 e",
"the.MAX.108 g",
"the.MAX.107 g",
"the.MAX.106 ErrNotFound",
"the.MAX.105 d",
"the.MAX.104 d",
"the.MAX.104 d",
"the.MAX.103 a",
"the.MAX.102 a",
"the.MAX.101 a",
"the.MAX.100 ErrNotFound",
"zzz.MAX.105 ErrNotFound",
},
},
{
description: "triple-0: three level-0 tables",
tables: []testTable{
{
level: 0,
fileNum: 10,
data: []string{
"the.SET.101 a",
"cat.SET.102 b",
"on_.SET.103 c",
"the.SET.104 d",
"mat.SET.105 e",
"the.DEL.106 ",
"the.SET.107 g",
},
},
{
level: 0,
fileNum: 11,
data: []string{
"awk.SET.111 w",
"cat.SET.112 x",
"man.SET.113 y",
"sed.SET.114 z",
},
},
{
level: 0,
fileNum: 12,
data: []string{
"the.DEL.121 ",
"cat.DEL.122 ",
"man.DEL.123 ",
"was.SET.124 D",
"not.SET.125 E",
"the.SET.126 F",
"man.SET.127 G",
},
},
},
queries: []string{
"aaa.MAX.105 ErrNotFound",
"awk.MAX.135 w",
"awk.MAX.125 w",
"awk.MAX.115 w",
"awk.MAX.105 ErrNotFound",
"cat.MAX.135 ErrNotFound",
"cat.MAX.125 ErrNotFound",
"cat.MAX.115 x",
"cat.MAX.105 b",
"man.MAX.135 G",
"man.MAX.125 ErrNotFound",
"man.MAX.115 y",
"man.MAX.105 ErrNotFound",
"on_.MAX.135 c",
"on_.MAX.125 c",
"on_.MAX.115 c",
"on_.MAX.105 c",
"the.MAX.135 F",
"the.MAX.127 F",
"the.MAX.126 F",
"the.MAX.125 ErrNotFound",
"the.MAX.122 ErrNotFound",
"the.MAX.121 ErrNotFound",
"the.MAX.120 g",
"the.MAX.115 g",
"the.MAX.114 g",
"the.MAX.111 g",
"the.MAX.110 g",
"the.MAX.108 g",
"the.MAX.107 g",
"the.MAX.106 ErrNotFound",
"the.MAX.105 d",
"the.MAX.104 d",
"the.MAX.104 d",
"the.MAX.103 a",
"the.MAX.102 a",
"the.MAX.101 a",
"the.MAX.100 ErrNotFound",
"zzz.MAX.105 ErrNotFound",
},
},
{
description: "quad-4: four level-4 tables",
tables: []testTable{
{
level: 4,
fileNum: 11,
data: []string{
"aardvark.SET.101 a1",
"alpaca__.SET.201 a2",
"anteater.SET.301 a3",
},
},
{
level: 4,
fileNum: 22,
data: []string{
"baboon__.SET.102 b1",
"baboon__.DEL.202 ",
"baboon__.SET.302 b3",
"bear____.SET.402 b4",
"bear____.DEL.502 ",
"buffalo_.SET.602 b6",
},
},
{
level: 4,
fileNum: 33,
data: []string{
"buffalo_.SET.103 B1",
},
},
{
level: 4,
fileNum: 44,
data: []string{
"chipmunk.SET.104 c1",
"chipmunk.SET.204 c2",
},
},
},
queries: []string{
"a_______.MAX.999 ErrNotFound",
"aardvark.MAX.999 a1",
"aardvark.MAX.102 a1",
"aardvark.MAX.101 a1",
"aardvark.MAX.100 ErrNotFound",
"alpaca__.MAX.999 a2",
"alpaca__.MAX.200 ErrNotFound",
"anteater.MAX.999 a3",
"anteater.MAX.302 a3",
"anteater.MAX.301 a3",
"anteater.MAX.300 ErrNotFound",
"anteater.MAX.000 ErrNotFound",
"b_______.MAX.999 ErrNotFound",
"baboon__.MAX.999 b3",
"baboon__.MAX.302 b3",
"baboon__.MAX.301 ErrNotFound",
"baboon__.MAX.202 ErrNotFound",
"baboon__.MAX.201 b1",
"baboon__.MAX.102 b1",
"baboon__.MAX.101 ErrNotFound",
"bear____.MAX.999 ErrNotFound",
"bear____.MAX.500 b4",
"bear____.MAX.000 ErrNotFound",
"buffalo_.MAX.999 b6",
"buffalo_.MAX.603 b6",
"buffalo_.MAX.602 b6",
"buffalo_.MAX.601 B1",
"buffalo_.MAX.104 B1",
"buffalo_.MAX.103 B1",
"buffalo_.MAX.102 ErrNotFound",
"buffalo_.MAX.000 ErrNotFound",
"c_______.MAX.999 ErrNotFound",
"chipmunk.MAX.999 c2",
"chipmunk.MAX.205 c2",
"chipmunk.MAX.204 c2",
"chipmunk.MAX.203 c1",
"chipmunk.MAX.105 c1",
"chipmunk.MAX.104 c1",
"chipmunk.MAX.103 ErrNotFound",
"chipmunk.MAX.000 ErrNotFound",
"d_______.MAX.999 ErrNotFound",
},
},
{
description: "complex: many tables at many levels",
tables: []testTable{
{
level: 0,
fileNum: 50,
data: []string{
"alfalfa__.SET.501 p1",
"asparagus.SET.502 p2",
"cabbage__.DEL.503 ",
},
},
{
level: 0,
fileNum: 51,
data: []string{
"asparagus.SET.511 q1",
"asparagus.SET.512 q2",
"asparagus.SET.513 q3",
"beans____.SET.514 q4",
"broccoli_.SET.515 q5",
"cabbage__.SET.516 q6",
"celery___.SET.517 q7",
},
},
{
level: 1,
fileNum: 40,
data: []string{
"alfalfa__.SET.410 r1",
"asparagus.SET.420 r2",
"arugula__.SET.430 r3",
},
},
{
level: 1,
fileNum: 41,
data: []string{
"beans____.SET.411 s1",
"beans____.SET.421 s2",
"bokchoy__.DEL.431 ",
"broccoli_.SET.441 s4",
},
},
{
level: 1,
fileNum: 42,
data: []string{
"cabbage__.SET.412 t1",
"corn_____.DEL.422 ",
},
},
{
level: 2,
fileNum: 30,
data: []string{
"alfalfa__.SET.310 u1",
"bokchoy__.SET.320 u2",
"celery___.SET.330 u3",
"corn_____.SET.340 u4",
},
},
},
queries: []string{
"a________.MAX.999 ErrNotFound",
"alfalfa__.MAX.520 p1",
"alfalfa__.MAX.510 p1",
"alfalfa__.MAX.500 r1",
"alfalfa__.MAX.400 u1",
"alfalfa__.MAX.300 ErrNotFound",
"asparagus.MAX.520 q3",
"asparagus.MAX.510 p2",
"asparagus.MAX.500 r2",
"asparagus.MAX.400 ErrNotFound",
"asparagus.MAX.300 ErrNotFound",
"arugula__.MAX.520 r3",
"arugula__.MAX.510 r3",
"arugula__.MAX.500 r3",
"arugula__.MAX.400 ErrNotFound",
"arugula__.MAX.300 ErrNotFound",
"beans____.MAX.520 q4",
"beans____.MAX.510 s2",
"beans____.MAX.500 s2",
"beans____.MAX.400 ErrNotFound",
"beans____.MAX.300 ErrNotFound",
"bokchoy__.MAX.520 ErrNotFound",
"bokchoy__.MAX.510 ErrNotFound",
"bokchoy__.MAX.500 ErrNotFound",
"bokchoy__.MAX.400 u2",
"bokchoy__.MAX.300 ErrNotFound",
"broccoli_.MAX.520 q5",
"broccoli_.MAX.510 s4",
"broccoli_.MAX.500 s4",
"broccoli_.MAX.400 ErrNotFound",
"broccoli_.MAX.300 ErrNotFound",
"cabbage__.MAX.520 q6",
"cabbage__.MAX.510 ErrNotFound",
"cabbage__.MAX.500 t1",
"cabbage__.MAX.400 ErrNotFound",
"cabbage__.MAX.300 ErrNotFound",
"celery___.MAX.520 q7",
"celery___.MAX.510 u3",
"celery___.MAX.500 u3",
"celery___.MAX.400 u3",
"celery___.MAX.300 ErrNotFound",
"corn_____.MAX.520 ErrNotFound",
"corn_____.MAX.510 ErrNotFound",
"corn_____.MAX.500 ErrNotFound",
"corn_____.MAX.400 u4",
"corn_____.MAX.300 ErrNotFound",
"d________.MAX.999 ErrNotFound",
},
},
{
description: "broken invariants 0: non-increasing level 0 file numbers",
badOrdering: true,
tables: []testTable{
{
level: 0,
fileNum: 20,
},
{
level: 0,
fileNum: 19,
},
},
},
{
description: "broken invariants 1: non-increasing level 0 file numbers",
badOrdering: true,
tables: []testTable{
{
level: 0,
fileNum: 20,
},
{
level: 0,
fileNum: 21,
},
{
level: 0,
fileNum: 21,
},
{
level: 0,
fileNum: 22,
},
},
},
{
description: "broken invariants 2: level non-0 overlapping internal key ranges",
badOrdering: true,
tables: []testTable{
{
level: 5,
fileNum: 11,
data: []string{
"bat.SET.101 xxx",
"dog.SET.102 xxx",
},
},
{
level: 5,
fileNum: 12,
data: []string{
"cow.SET.103 xxx",
"pig.SET.104 xxx",
},
},
},
},
}
icmp := internalKeyComparer{db.DefaultComparer}
for _, tc := range testCases {
desc := tc.description[:strings.Index(tc.description, ":")]
// m is a map from file numbers to DBs.
m := map[uint64]db.DB{}
tiFinder := tableIkeyFinderFunc(func(fileNum uint64, ikey internalKey) (db.Iterator, error) {
d, ok := m[fileNum]
if !ok {
return nil, errors.New("no such file")
}
return d.Find(ikey, nil), nil
})
v := version{}
for _, tt := range tc.tables {
d := memdb.New(&db.Options{
Comparer: icmp,
})
defer d.Close()
m[tt.fileNum] = d
var smallest, largest internalKey
for i, datum := range tt.data {
s := strings.Split(datum, " ")
ikey := makeIkey(s[0])
err := d.Set(ikey, []byte(s[1]), nil)
if err != nil {
t.Fatalf("desc=%q: memdb Set: %v", desc, err)
}
if i == 0 {
smallest = ikey
largest = ikey
} else {
if icmp.Compare(ikey, smallest) < 0 {
smallest = ikey
}
if icmp.Compare(ikey, largest) > 0 {
largest = ikey
}
}
}
v.files[tt.level] = append(v.files[tt.level], fileMetadata{
fileNum: tt.fileNum,
smallest: smallest,
largest: largest,
})
}
err := v.checkOrdering(icmp)
if tc.badOrdering && err == nil {
t.Errorf("desc=%q: want bad ordering, got nil error", desc)
continue
} else if !tc.badOrdering && err != nil {
t.Errorf("desc=%q: bad ordering: %v", desc, err)
continue
}
for _, query := range tc.queries {
s := strings.Split(query, " ")
value, err := v.get(makeIkey(s[0]), tiFinder, db.DefaultComparer, nil)
got, want := "", s[1]
if err != nil {
if err != db.ErrNotFound {
t.Errorf("desc=%q: query=%q: %v", desc, s[0], err)
continue
}
got = "ErrNotFound"
} else {
got = string(value)
}
if got != want {
t.Errorf("desc=%q: query=%q: got %q, want %q", desc, s[0], got, want)
}
}
}
}