func BenchmarkSstRekey(b *testing.B) {
// TODO(dan): DRY this with BenchmarkRocksDBSstFileReader.
dir, cleanupFn := testutils.TempDir(b, 1)
defer cleanupFn()
sstPath := filepath.Join(dir, "sst")
{
const maxEntries = 100000
const keyLen = 10
const valLen = 100
b.SetBytes(keyLen + valLen)
ts := hlc.Timestamp{WallTime: timeutil.Now().UnixNano()}
kv := engine.MVCCKeyValue{
Key: engine.MVCCKey{Key: roachpb.Key(make([]byte, keyLen)), Timestamp: ts},
Value: make([]byte, valLen),
}
sst := engine.MakeRocksDBSstFileWriter()
if err := sst.Open(sstPath); err != nil {
b.Fatal(sst)
}
var entries = b.N
if entries > maxEntries {
entries = maxEntries
}
for i := 0; i < entries; i++ {
payload := []byte(fmt.Sprintf("%09d", i))
kv.Key.Key = kv.Key.Key[:0]
kv.Key.Key = encoding.EncodeUvarintAscending(kv.Key.Key, uint64(i)) // tableID
kv.Key.Key = encoding.EncodeUvarintAscending(kv.Key.Key, 0) // indexID
kv.Key.Key = encoding.EncodeBytesAscending(kv.Key.Key, payload)
kv.Key.Key = keys.MakeRowSentinelKey(kv.Key.Key)
copy(kv.Value, payload)
if err := sst.Add(kv); err != nil {
b.Fatal(err)
}
}
if err := sst.Close(); err != nil {
b.Fatal(err)
}
}
const newTableID = 100
b.ResetTimer()
sst, err := engine.MakeRocksDBSstFileReader()
if err != nil {
b.Fatal(err)
}
if err := sst.AddFile(sstPath); err != nil {
b.Fatal(err)
}
defer sst.Close()
count := 0
iterateFn := sql.MakeRekeyMVCCKeyValFunc(newTableID, func(kv engine.MVCCKeyValue) (bool, error) {
count++
if count >= b.N {
return true, nil
}
return false, nil
})
for {
if err := sst.Iterate(engine.MVCCKey{Key: keys.MinKey}, engine.MVCCKey{Key: keys.MaxKey}, iterateFn); err != nil {
b.Fatal(err)
}
if count >= b.N {
break
}
}
}
// Backup exports a snapshot of every kv entry into ranged sstables.
//
// The output is an sstable per range with files in the following locations:
// - /<base>/<node_id>/<key_range>/data.sst
// - <base> is given by the user and is expected to eventually be cloud storage
// - The <key_range>s are non-overlapping.
//
// TODO(dan): Bikeshed this directory structure and naming.
func Backup(
ctx context.Context, db client.DB, base string, endTime hlc.Timestamp,
) (desc sqlbase.BackupDescriptor, retErr error) {
// TODO(dan): Optionally take a start time for an incremental backup.
// TODO(dan): Take a uri for the path prefix and support various cloud storages.
// TODO(dan): Figure out how permissions should work. #6713 is tracking this
// for grpc.
var rangeDescs []roachpb.RangeDescriptor
var sqlDescs []sqlbase.Descriptor
opt := client.TxnExecOptions{
AutoRetry: true,
AutoCommit: true,
}
{
// TODO(dan): Pick an appropriate end time and set it in the txn.
txn := client.NewTxn(ctx, db)
err := txn.Exec(opt, func(txn *client.Txn, opt *client.TxnExecOptions) error {
var err error
SetTxnTimestamps(txn, endTime)
rangeDescs, err = AllRangeDescriptors(txn)
if err != nil {
return err
}
sqlDescs, err = allSQLDescriptors(txn)
return err
})
if err != nil {
return sqlbase.BackupDescriptor{}, err
}
}
var dataSize int64
backupDescs := make([]sqlbase.BackupRangeDescriptor, len(rangeDescs))
crc := crc32.New(crc32.MakeTable(crc32.Castagnoli))
for i, rangeDesc := range rangeDescs {
backupDescs[i] = sqlbase.BackupRangeDescriptor{
StartKey: rangeDesc.StartKey.AsRawKey(),
EndKey: rangeDesc.EndKey.AsRawKey(),
StartTime: hlc.Timestamp{},
}
if backupDescs[i].StartKey.Compare(keys.LocalMax) < 0 {
backupDescs[i].StartKey = keys.LocalMax
}
nodeID := 0
dir := filepath.Join(base, fmt.Sprintf("%03d", nodeID))
dir = filepath.Join(dir, fmt.Sprintf("%x-%x", rangeDesc.StartKey, rangeDesc.EndKey))
if err := os.MkdirAll(dir, 0700); err != nil {
return sqlbase.BackupDescriptor{}, err
}
var kvs []client.KeyValue
txn := client.NewTxn(ctx, db)
err := txn.Exec(opt, func(txn *client.Txn, opt *client.TxnExecOptions) error {
var err error
SetTxnTimestamps(txn, endTime)
// TODO(dan): Iterate with some batch size.
kvs, err = txn.Scan(backupDescs[i].StartKey, backupDescs[i].EndKey, 0)
return err
})
if err != nil {
return sqlbase.BackupDescriptor{}, err
}
if len(kvs) == 0 {
if log.V(1) {
log.Infof(ctx, "skipping backup of empty range %s-%s",
backupDescs[i].StartKey, backupDescs[i].EndKey)
}
continue
}
backupDescs[i].Path = filepath.Join(dir, dataSSTableName)
writeSST := func() (writeSSTErr error) {
// This is a function so the defered Close (and resultant flush) is
// called before the checksum is computed.
sst := engine.MakeRocksDBSstFileWriter()
if err := sst.Open(backupDescs[i].Path); err != nil {
return err
}
defer func() {
if closeErr := sst.Close(); closeErr != nil && writeSSTErr == nil {
writeSSTErr = closeErr
}
}()
// TODO(dan): Move all this iteration into cpp to avoid the cgo calls.
for _, kv := range kvs {
mvccKV := engine.MVCCKeyValue{
Key: engine.MVCCKey{Key: kv.Key, Timestamp: kv.Value.Timestamp},
Value: kv.Value.RawBytes,
}
if err := sst.Add(mvccKV); err != nil {
return err
}
}
dataSize += sst.DataSize
return nil
}
if err := writeSST(); err != nil {
return sqlbase.BackupDescriptor{}, err
}
crc.Reset()
f, err := os.Open(backupDescs[i].Path)
if err != nil {
return sqlbase.BackupDescriptor{}, err
}
defer f.Close()
if _, err := io.Copy(crc, f); err != nil {
return sqlbase.BackupDescriptor{}, err
}
backupDescs[i].CRC = crc.Sum32()
}
desc = sqlbase.BackupDescriptor{
EndTime: endTime,
Ranges: backupDescs,
SQL: sqlDescs,
DataSize: dataSize,
}
descBuf, err := desc.Marshal()
if err != nil {
return sqlbase.BackupDescriptor{}, err
}
if err = ioutil.WriteFile(filepath.Join(base, backupDescriptorName), descBuf, 0600); err != nil {
return sqlbase.BackupDescriptor{}, err
}
return desc, nil
}
