// fillRange writes keys with the given prefix and associated values
// until bytes bytes have been written or the given range has split.
func fillRange(store *storage.Store, rangeID roachpb.RangeID, prefix roachpb.Key, bytes int64, t *testing.T) {
src := rand.New(rand.NewSource(0))
for {
var ms engine.MVCCStats
if err := engine.MVCCGetRangeStats(store.Engine(), rangeID, &ms); err != nil {
t.Fatal(err)
}
keyBytes, valBytes := ms.KeyBytes, ms.ValBytes
if keyBytes+valBytes >= bytes {
return
}
key := append(append([]byte(nil), prefix...), randutil.RandBytes(src, 100)...)
key = keys.MakeNonColumnKey(key)
val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
pArgs := putArgs(key, val)
_, err := client.SendWrappedWith(store, nil, roachpb.Header{
RangeID: rangeID,
}, &pArgs)
// When the split occurs in the background, our writes may start failing.
// We know we can stop writing when this happens.
if _, ok := err.(*roachpb.RangeKeyMismatchError); ok {
return
} else if err != nil {
t.Fatal(err)
}
}
}
func fillTestRange(t testing.TB, rep *Replica, size int) {
src := rand.New(rand.NewSource(0))
for i := 0; i < snapSize/(keySize+valSize); i++ {
key := keys.MakeRowSentinelKey(randutil.RandBytes(src, keySize))
val := randutil.RandBytes(src, valSize)
pArgs := putArgs(key, val)
if _, pErr := client.SendWrappedWith(rep, nil, roachpb.Header{
RangeID: rangeID,
}, &pArgs); pErr != nil {
t.Fatal(pErr)
}
}
}
func writeRandomDataToRange(t testing.TB, store *storage.Store, rangeID roachpb.RangeID, keyPrefix []byte) {
src := rand.New(rand.NewSource(0))
for i := 0; i < 100; i++ {
key := append([]byte(nil), keyPrefix...)
key = append(key, randutil.RandBytes(src, int(src.Int31n(1<<7)))...)
key = keys.MakeNonColumnKey(key)
val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
pArgs := putArgs(key, val)
if _, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
RangeID: rangeID,
}, &pArgs); pErr != nil {
t.Fatal(pErr)
}
}
}
func runMVCCConditionalPut(emk engineMaker, valueSize int, createFirst bool, b *testing.B) {
rng, _ := randutil.NewPseudoRand()
value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
eng, stopper := emk(b, fmt.Sprintf("cput_%d", valueSize))
defer stopper.Stop()
b.SetBytes(int64(valueSize))
var expected *roachpb.Value
if createFirst {
for i := 0; i < b.N; i++ {
key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
ts := makeTS(timeutil.Now().UnixNano(), 0)
if err := MVCCPut(context.Background(), eng, nil, key, ts, value, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
expected = &value
}
b.ResetTimer()
for i := 0; i < b.N; i++ {
key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
ts := makeTS(timeutil.Now().UnixNano(), 0)
if err := MVCCConditionalPut(context.Background(), eng, nil, key, ts, value, expected, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
b.StopTimer()
}
func runMVCCConditionalPut(valueSize int, createFirst bool, b *testing.B) {
rng, _ := randutil.NewPseudoRand()
value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
stopper := stop.NewStopper()
defer stopper.Stop()
rocksdb := NewInMem(roachpb.Attributes{}, testCacheSize, stopper)
b.SetBytes(int64(valueSize))
var expected *roachpb.Value
if createFirst {
for i := 0; i < b.N; i++ {
key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
ts := makeTS(timeutil.Now().UnixNano(), 0)
if err := MVCCPut(rocksdb, nil, key, ts, value, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
expected = &value
}
b.ResetTimer()
for i := 0; i < b.N; i++ {
key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(i)))
ts := makeTS(timeutil.Now().UnixNano(), 0)
if err := MVCCConditionalPut(rocksdb, nil, key, ts, value, expected, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
b.StopTimer()
}
func BenchmarkMVCCPutDelete_RocksDB(b *testing.B) {
const cacheSize = 1 << 30 // 1 GB
rocksdb, stopper := setupMVCCInMemRocksDB(b, "put_delete")
defer stopper.Stop()
r := rand.New(rand.NewSource(int64(timeutil.Now().UnixNano())))
value := roachpb.MakeValueFromBytes(randutil.RandBytes(r, 10))
zeroTS := roachpb.ZeroTimestamp
var blockNum int64
for i := 0; i < b.N; i++ {
blockID := r.Int63()
blockNum++
key := encoding.EncodeVarintAscending(nil, blockID)
key = encoding.EncodeVarintAscending(key, blockNum)
if err := MVCCPut(context.Background(), rocksdb, nil, key, zeroTS, value, nil /* txn */); err != nil {
b.Fatal(err)
}
if err := MVCCDelete(context.Background(), rocksdb, nil, key, zeroTS, nil /* txn */); err != nil {
b.Fatal(err)
}
}
}
func TestApproximateSize(t *testing.T) {
defer leaktest.AfterTest(t)
runWithAllEngines(func(engine Engine, t *testing.T) {
var (
count = 10000
keys = make([]proto.EncodedKey, count)
values = make([][]byte, count) // Random values to prevent compression
rand, _ = randutil.NewPseudoRand()
valueLen = 10
)
for i := 0; i < count; i++ {
keys[i] = []byte(fmt.Sprintf("key%8d", i))
values[i] = randutil.RandBytes(rand, valueLen)
}
insertKeysAndValues(keys, values, engine, t)
if err := engine.Flush(); err != nil {
t.Fatalf("Error flushing InMem: %s", err)
}
sizePerRecord := (len([]byte(keys[0])) + valueLen)
verifyApproximateSize(keys, engine, sizePerRecord, 0.15, t)
verifyApproximateSize(keys[:count/2], engine, sizePerRecord, 0.15, t)
verifyApproximateSize(keys[:count/4], engine, sizePerRecord, 0.15, t)
}, t)
}
func BenchmarkMVCCPutDelete(b *testing.B) {
const cacheSize = 1 << 30 // 1 GB
stopper := stop.NewStopper()
rocksdb := NewInMem(roachpb.Attributes{}, cacheSize, stopper)
defer stopper.Stop()
r := rand.New(rand.NewSource(int64(timeutil.Now().UnixNano())))
value := roachpb.MakeValueFromBytes(randutil.RandBytes(r, 10))
zeroTS := roachpb.ZeroTimestamp
var blockNum int64
for i := 0; i < b.N; i++ {
blockID := r.Int63()
blockNum++
key := encoding.EncodeVarintAscending(nil, blockID)
key = encoding.EncodeVarintAscending(key, blockNum)
if err := MVCCPut(rocksdb, nil, key, zeroTS, value, nil /* txn */); err != nil {
b.Fatal(err)
}
if err := MVCCDelete(rocksdb, nil, key, zeroTS, nil /* txn */); err != nil {
b.Fatal(err)
}
}
}
func TestRandBytes(t *testing.T) {
rand, _ := randutil.NewPseudoRand()
for i := 0; i < 100; i++ {
x := randutil.RandBytes(rand, i)
if len(x) != i {
t.Errorf("got array with unexpected length: %d (expected %d)", len(x), i)
}
}
}
// fillRange writes keys with the given prefix and associated values
// until bytes bytes have been written.
func fillRange(store *storage.Store, rangeID roachpb.RangeID, prefix roachpb.Key, bytes int64, t *testing.T) {
src := rand.New(rand.NewSource(0))
for {
var ms engine.MVCCStats
if err := engine.MVCCGetRangeStats(store.Engine(), rangeID, &ms); err != nil {
t.Fatal(err)
}
keyBytes, valBytes := ms.KeyBytes, ms.ValBytes
if keyBytes+valBytes >= bytes {
return
}
key := append(append([]byte(nil), prefix...), randutil.RandBytes(src, 100)...)
val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
pArgs := putArgs(key, val, rangeID, store.StoreID())
if _, err := client.SendWrapped(store, nil, &pArgs); err != nil {
t.Fatal(err)
}
}
}
// startTestWriter creates a writer which initiates a sequence of
// transactions, each which writes up to 10 times to random keys with
// random values. If not nil, txnChannel is written to non-blockingly
// every time a new transaction starts.
func startTestWriter(db *client.DB, i int64, valBytes int32, wg *sync.WaitGroup, retries *int32,
txnChannel chan struct{}, done <-chan struct{}, t *testing.T) {
src := rand.New(rand.NewSource(i))
defer func() {
if wg != nil {
wg.Done()
}
}()
for j := 0; ; j++ {
select {
case <-done:
return
default:
first := true
err := db.Txn(func(txn *client.Txn) error {
if first && txnChannel != nil {
select {
case txnChannel <- struct{}{}:
default:
}
} else if !first && retries != nil {
atomic.AddInt32(retries, 1)
}
first = false
for j := 0; j <= int(src.Int31n(10)); j++ {
key := randutil.RandBytes(src, 10)
val := randutil.RandBytes(src, int(src.Int31n(valBytes)))
if err := txn.Put(key, val); err != nil {
log.Infof("experienced an error in routine %d: %s", i, err)
return err
}
}
return nil
})
if err != nil {
t.Error(err)
} else {
time.Sleep(1 * time.Millisecond)
}
}
}
}
func fillTestRange(t testing.TB, rep *Replica, size int64) {
src := rand.New(rand.NewSource(0))
for i := int64(0); i < size/int64(keySize+valSize); i++ {
key := keys.MakeRowSentinelKey(randutil.RandBytes(src, keySize))
val := randutil.RandBytes(src, valSize)
pArgs := putArgs(key, val)
if _, pErr := client.SendWrappedWith(rep, nil, roachpb.Header{
RangeID: rangeID,
}, &pArgs); pErr != nil {
t.Fatal(pErr)
}
}
rep.mu.Lock()
after := rep.mu.state.Stats.Total()
rep.mu.Unlock()
if after < size {
t.Fatalf("range not full after filling: wrote %d, but range at %d", size, after)
}
}
// fillRange writes keys with the given prefix and associated values
// until bytes bytes have been written.
func fillRange(store *storage.Store, rangeID proto.RangeID, prefix proto.Key, bytes int64, t *testing.T) {
src := rand.New(rand.NewSource(0))
for {
var ms engine.MVCCStats
if err := engine.MVCCGetRangeStats(store.Engine(), rangeID, &ms); err != nil {
t.Fatal(err)
}
keyBytes, valBytes := ms.KeyBytes, ms.ValBytes
if keyBytes+valBytes >= bytes {
return
}
key := append(append([]byte(nil), prefix...), randutil.RandBytes(src, 100)...)
val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
pArgs := putArgs(key, val, rangeID, store.StoreID())
pArgs.Timestamp = store.Clock().Now()
if _, err := store.ExecuteCmd(context.Background(), &pArgs); err != nil {
t.Fatal(err)
}
}
}
// setupMVCCData writes up to numVersions values at each of numKeys
// keys. The number of versions written for each key is chosen
// randomly according to a uniform distribution. Each successive
// version is written starting at 5ns and then in 5ns increments. This
// allows scans at various times, starting at t=5ns, and continuing to
// t=5ns*(numVersions+1). A version for each key will be read on every
// such scan, but the dynamics of the scan will change depending on
// the historical timestamp. Earlier timestamps mean scans which must
// skip more historical versions; later timestamps mean scans which
// skip fewer.
//
// The creation of the rocksdb database is time consuming, especially
// for larger numbers of versions. The database is persisted between
// runs and stored in the current directory as
// "mvcc_scan_<versions>_<keys>_<valueBytes>".
func setupMVCCScanData(numVersions, numKeys, valueBytes int, b *testing.B) (*RocksDB, *stop.Stopper) {
loc := fmt.Sprintf("mvcc_scan_%d_%d_%d", numVersions, numKeys, valueBytes)
exists := true
if _, err := os.Stat(loc); os.IsNotExist(err) {
exists = false
}
const cacheSize = 8 << 30 // 8 GB
stopper := stop.NewStopper()
rocksdb := NewRocksDB(roachpb.Attributes{}, loc, cacheSize, stopper)
if err := rocksdb.Open(); err != nil {
b.Fatalf("could not create new rocksdb db instance at %s: %v", loc, err)
}
if exists {
return rocksdb, stopper
}
log.Infof("creating mvcc data: %s", loc)
rng, _ := randutil.NewPseudoRand()
keys := make([]roachpb.Key, numKeys)
nvs := make([]int, numKeys)
for t := 1; t <= numVersions; t++ {
walltime := int64(5 * t)
ts := makeTS(walltime, 0)
batch := rocksdb.NewBatch()
for i := 0; i < numKeys; i++ {
if t == 1 {
keys[i] = roachpb.Key(encoding.EncodeUvarint([]byte("key-"), uint64(i)))
nvs[i] = rand.Intn(numVersions) + 1
}
// Only write values if this iteration is less than the random
// number of versions chosen for this key.
if t <= nvs[i] {
value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueBytes))
value.InitChecksum(keys[i])
if err := MVCCPut(batch, nil, keys[i], ts, value, nil); err != nil {
b.Fatal(err)
}
}
}
if err := batch.Commit(); err != nil {
b.Fatal(err)
}
batch.Close()
}
rocksdb.CompactRange(nil, nil)
return rocksdb, stopper
}
// TestPut starts up an N node cluster and runs N workers that write
// to independent keys.
func TestPut(t *testing.T) {
l := localcluster.Create(*numNodes, stopper)
l.Start()
defer l.Stop()
db, dbStopper := makeDBClient(t, l, 0)
defer dbStopper.Stop()
if err := configutil.SetDefaultRangeMaxBytes(db, *rangeMaxBytes); err != nil {
t.Fatal(err)
}
checkRangeReplication(t, l, 20*time.Second)
errs := make(chan error, *numNodes)
start := time.Now()
deadline := start.Add(*duration)
var count int64
for i := 0; i < *numNodes; i++ {
go func() {
r, _ := randutil.NewPseudoRand()
value := randutil.RandBytes(r, 8192)
for time.Now().Before(deadline) {
k := atomic.AddInt64(&count, 1)
v := value[:r.Intn(len(value))]
if err := db.Put(fmt.Sprintf("%08d", k), v); err != nil {
errs <- err
return
}
}
errs <- nil
}()
}
for i := 0; i < *numNodes; {
select {
case <-stopper:
t.Fatalf("interrupted")
case err := <-errs:
if err != nil {
t.Fatal(err)
}
i++
case <-time.After(1 * time.Second):
// Periodically print out progress so that we know the test is still
// running.
log.Infof("%d", atomic.LoadInt64(&count))
}
}
elapsed := time.Since(start)
log.Infof("%d %.1f/sec", count, float64(count)/elapsed.Seconds())
}
func BenchmarkPeekLengthBytesDescending(b *testing.B) {
rng, _ := randutil.NewPseudoRand()
vals := make([][]byte, 10000)
for i := range vals {
vals[i] = EncodeBytesDescending(nil, randutil.RandBytes(rng, 100))
}
b.ResetTimer()
for i := 0; i < b.N; i++ {
_, _ = PeekLength(vals[i%len(vals)])
}
}
// TestPut starts up an N node cluster and runs N workers that write
// to independent keys.
func TestPut(t *testing.T) {
c := StartCluster(t)
defer c.AssertAndStop(t)
db, dbStopper := makeClient(t, c.ConnString(0))
defer dbStopper.Stop()
errs := make(chan error, c.NumNodes())
start := time.Now()
deadline := start.Add(*flagDuration)
var count int64
for i := 0; i < c.NumNodes(); i++ {
go func() {
r, _ := randutil.NewPseudoRand()
value := randutil.RandBytes(r, 8192)
for time.Now().Before(deadline) {
k := atomic.AddInt64(&count, 1)
v := value[:r.Intn(len(value))]
if pErr := db.Put(fmt.Sprintf("%08d", k), v); pErr != nil {
errs <- pErr.GoError()
return
}
}
errs <- nil
}()
}
for i := 0; i < c.NumNodes(); {
baseCount := atomic.LoadInt64(&count)
select {
case <-stopper:
t.Fatalf("interrupted")
case err := <-errs:
if err != nil {
t.Fatal(err)
}
i++
case <-time.After(1 * time.Second):
// Periodically print out progress so that we know the test is still
// running.
count := atomic.LoadInt64(&count)
log.Infof("%d (%d/s)", count, count-baseCount)
c.Assert(t)
}
}
elapsed := time.Since(start)
log.Infof("%d %.1f/sec", count, float64(count)/elapsed.Seconds())
}
func BenchmarkReplicaSnapshot(b *testing.B) {
defer tracing.Disable()()
defer config.TestingDisableTableSplits()()
store, stopper, _ := createTestStore(b)
// We want to manually control the size of the raft log.
store.DisableRaftLogQueue(true)
defer stopper.Stop()
const rangeID = 1
const keySize = 1 << 7 // 128 B
const valSize = 1 << 10 // 1 KiB
const snapSize = 1 << 25 // 32 MiB
rep, err := store.GetReplica(rangeID)
if err != nil {
b.Fatal(err)
}
src := rand.New(rand.NewSource(0))
for i := 0; i < snapSize/(keySize+valSize); i++ {
key := keys.MakeRowSentinelKey(randutil.RandBytes(src, keySize))
val := randutil.RandBytes(src, valSize)
pArgs := putArgs(key, val)
if _, pErr := client.SendWrappedWith(rep, nil, roachpb.Header{
RangeID: rangeID,
}, &pArgs); pErr != nil {
b.Fatal(pErr)
}
}
b.ResetTimer()
for i := 0; i < b.N; i++ {
if _, err := rep.GetSnapshot(); err != nil {
b.Fatal(err)
}
}
}
func BenchmarkDecodeStringDescending(b *testing.B) {
rng, _ := randutil.NewPseudoRand()
vals := make([][]byte, 10000)
for i := range vals {
vals[i] = EncodeStringDescending(nil, string(randutil.RandBytes(rng, 100)))
}
buf := make([]byte, 0, 1000)
b.ResetTimer()
for i := 0; i < b.N; i++ {
_, _, _ = DecodeStringDescending(vals[i%len(vals)], buf)
}
}
func BenchmarkEncodeString(b *testing.B) {
rng, _ := randutil.NewPseudoRand()
vals := make([]string, 10000)
for i := range vals {
vals[i] = string(randutil.RandBytes(rng, 100))
}
buf := make([]byte, 0, 1000)
b.ResetTimer()
for i := 0; i < b.N; i++ {
_ = EncodeStringAscending(buf, vals[i%len(vals)])
}
}
func BenchmarkDecodeKeyBytes(b *testing.B) {
rng, _ := randutil.NewPseudoRand()
vals := make([][]byte, 10000)
for i := range vals {
vals[i] = EncodeBytes(nil, randutil.RandBytes(rng, 100))
}
result := []byte(nil)
b.ResetTimer()
for i := 0; i < b.N; i++ {
_ = DecodeKey(vals[i%len(vals)], "%s", &result)
}
}
func writeRandomDataToRange(
t testing.TB,
store *storage.Store,
rangeID roachpb.RangeID,
keyPrefix []byte,
) (midpoint []byte) {
src := rand.New(rand.NewSource(0))
for i := 0; i < 100; i++ {
key := append([]byte(nil), keyPrefix...)
key = append(key, randutil.RandBytes(src, int(src.Int31n(1<<7)))...)
key = keys.MakeRowSentinelKey(key)
val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
pArgs := putArgs(key, val)
if _, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
RangeID: rangeID,
}, &pArgs); pErr != nil {
t.Fatal(pErr)
}
}
// Return approximate midway point ("Z" in string "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz").
midKey := append([]byte(nil), keyPrefix...)
midKey = append(midKey, []byte("Z")...)
return keys.MakeRowSentinelKey(midKey)
}
func BenchmarkEncodeKeyBytes(b *testing.B) {
rng, _ := randutil.NewPseudoRand()
vals := make([]interface{}, 10000)
for i := range vals {
vals[i] = randutil.RandBytes(rng, 100)
}
buf := make([]byte, 0, 1000)
b.ResetTimer()
for i := 0; i < b.N; i++ {
_ = EncodeKey(buf, "%s", vals[i%len(vals)])
}
}
func testPutInner(t *testing.T, c cluster.Cluster, cfg cluster.TestConfig) {
db, dbStopper := c.NewClient(t, 0)
defer dbStopper.Stop()
errs := make(chan error, c.NumNodes())
start := timeutil.Now()
deadline := start.Add(cfg.Duration)
var count int64
for i := 0; i < c.NumNodes(); i++ {
go func() {
r, _ := randutil.NewPseudoRand()
value := randutil.RandBytes(r, 8192)
for timeutil.Now().Before(deadline) {
k := atomic.AddInt64(&count, 1)
v := value[:r.Intn(len(value))]
if err := db.Put(fmt.Sprintf("%08d", k), v); err != nil {
errs <- err
return
}
}
errs <- nil
}()
}
for i := 0; i < c.NumNodes(); {
baseCount := atomic.LoadInt64(&count)
select {
case <-stopper:
t.Fatalf("interrupted")
case err := <-errs:
if err != nil {
t.Fatal(err)
}
i++
case <-time.After(1 * time.Second):
// Periodically print out progress so that we know the test is still
// running.
loadedCount := atomic.LoadInt64(&count)
log.Infof(context.Background(), "%d (%d/s)", loadedCount, loadedCount-baseCount)
c.Assert(t)
cluster.Consistent(t, c)
}
}
elapsed := timeutil.Since(start)
log.Infof(context.Background(), "%d %.1f/sec", count, float64(count)/elapsed.Seconds())
}
func runMVCCPut(valueSize int, b *testing.B) {
rng, _ := randutil.NewPseudoRand()
value := proto.Value{Bytes: randutil.RandBytes(rng, valueSize)}
keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
rocksdb := NewInMem(proto.Attributes{Attrs: []string{"ssd"}}, testCacheSize)
defer rocksdb.Close()
b.SetBytes(int64(valueSize))
b.ResetTimer()
for i := 0; i < b.N; i++ {
key := proto.Key(encoding.EncodeUvarint(keyBuf[0:4], uint64(i)))
ts := makeTS(time.Now().UnixNano(), 0)
if err := MVCCPut(rocksdb, nil, key, ts, value, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
b.StopTimer()
}
func runMVCCBatchPut(valueSize, batchSize int, b *testing.B) {
defer tracing.Disable()()
rng, _ := randutil.NewPseudoRand()
value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
stopper := stop.NewStopper()
defer stopper.Stop()
rocksdb := NewInMem(roachpb.Attributes{}, testCacheSize, stopper)
b.SetBytes(int64(valueSize))
b.ResetTimer()
for i := 0; i < b.N; i += batchSize {
end := i + batchSize
if end > b.N {
end = b.N
}
batch := rocksdb.NewBatch()
for j := i; j < end; j++ {
key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(j)))
ts := makeTS(time.Now().UnixNano(), 0)
if err := MVCCPut(batch, nil, key, ts, value, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
if err := batch.Commit(); err != nil {
b.Fatal(err)
}
batch.Close()
}
b.StopTimer()
}
func runMVCCBatchPut(emk engineMaker, valueSize, batchSize int, b *testing.B) {
rng, _ := randutil.NewPseudoRand()
value := roachpb.MakeValueFromBytes(randutil.RandBytes(rng, valueSize))
keyBuf := append(make([]byte, 0, 64), []byte("key-")...)
stopper := stop.NewStopper()
eng, stopper := emk(b, fmt.Sprintf("batch_put_%d_%d", valueSize, batchSize))
defer stopper.Stop()
b.SetBytes(int64(valueSize))
b.ResetTimer()
for i := 0; i < b.N; i += batchSize {
end := i + batchSize
if end > b.N {
end = b.N
}
batch := eng.NewBatch()
for j := i; j < end; j++ {
key := roachpb.Key(encoding.EncodeUvarintAscending(keyBuf[:4], uint64(j)))
ts := makeTS(timeutil.Now().UnixNano(), 0)
if err := MVCCPut(context.Background(), batch, nil, key, ts, value, nil); err != nil {
b.Fatalf("failed put: %s", err)
}
}
if err := batch.Commit(); err != nil {
b.Fatal(err)
}
batch.Close()
}
b.StopTimer()
}
// TestStoreRangeSplitStats starts by splitting the system keys from user-space
// keys and verifying that the user space side of the split (which is empty),
// has all zeros for stats. It then writes random data to the user space side,
// splits it halfway and verifies the two splits have stats exactly equaling
// the pre-split.
func TestStoreRangeSplitStats(t *testing.T) {
defer leaktest.AfterTest(t)
store, stopper := createTestStore(t)
defer stopper.Stop()
// Split the range after the last table data key.
keyPrefix := keys.MakeTablePrefix(keys.MaxReservedDescID + 1)
keyPrefix = keys.MakeNonColumnKey(keyPrefix)
args := adminSplitArgs(roachpb.KeyMin, keyPrefix)
if _, err := client.SendWrapped(rg1(store), nil, &args); err != nil {
t.Fatal(err)
}
// Verify empty range has empty stats.
rng := store.LookupReplica(keyPrefix, nil)
// NOTE that this value is expected to change over time, depending on what
// we store in the sys-local keyspace. Update it accordingly for this test.
if err := verifyRangeStats(store.Engine(), rng.Desc().RangeID, engine.MVCCStats{}); err != nil {
t.Fatal(err)
}
// Write random data.
src := rand.New(rand.NewSource(0))
for i := 0; i < 100; i++ {
key := append([]byte(nil), keyPrefix...)
key = append(key, randutil.RandBytes(src, int(src.Int31n(1<<7)))...)
key = keys.MakeNonColumnKey(key)
val := randutil.RandBytes(src, int(src.Int31n(1<<8)))
pArgs := putArgs(key, val)
if _, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
RangeID: rng.Desc().RangeID,
}, &pArgs); err != nil {
t.Fatal(err)
}
}
// Get the range stats now that we have data.
var ms engine.MVCCStats
if err := engine.MVCCGetRangeStats(store.Engine(), rng.Desc().RangeID, &ms); err != nil {
t.Fatal(err)
}
// Split the range at approximate halfway point ("Z" in string "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz").
midKey := append([]byte(nil), keyPrefix...)
midKey = append(midKey, []byte("Z")...)
midKey = keys.MakeNonColumnKey(midKey)
args = adminSplitArgs(keyPrefix, midKey)
if _, err := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
RangeID: rng.Desc().RangeID,
}, &args); err != nil {
t.Fatal(err)
}
var msLeft, msRight engine.MVCCStats
if err := engine.MVCCGetRangeStats(store.Engine(), rng.Desc().RangeID, &msLeft); err != nil {
t.Fatal(err)
}
rngRight := store.LookupReplica(midKey, nil)
if err := engine.MVCCGetRangeStats(store.Engine(), rngRight.Desc().RangeID, &msRight); err != nil {
t.Fatal(err)
}
// The stats should be exactly equal when added.
expMS := engine.MVCCStats{
LiveBytes: msLeft.LiveBytes + msRight.LiveBytes,
KeyBytes: msLeft.KeyBytes + msRight.KeyBytes,
ValBytes: msLeft.ValBytes + msRight.ValBytes,
IntentBytes: msLeft.IntentBytes + msRight.IntentBytes,
LiveCount: msLeft.LiveCount + msRight.LiveCount,
KeyCount: msLeft.KeyCount + msRight.KeyCount,
ValCount: msLeft.ValCount + msRight.ValCount,
IntentCount: msLeft.IntentCount + msRight.IntentCount,
}
ms.SysBytes, ms.SysCount = 0, 0
if !reflect.DeepEqual(expMS, ms) {
t.Errorf("expected left and right ranges to equal original: %+v + %+v != %+v", msLeft, msRight, ms)
}
}
func makeCommandID() string {
return string(randutil.RandBytes(testRand, commandIDLen))
}
// TestChaos starts up a cluster and, for each node, a worker writing to
// independent keys, while nodes are being killed and restarted continuously.
// The test measures not write performance, but cluster recovery.
func TestChaos(t *testing.T) {
t.Skip("TODO(tschottdorf): currently unstable")
l := localcluster.Create(*numNodes, stopper)
l.Start()
defer l.AssertAndStop(t)
checkRangeReplication(t, l, 20*time.Second)
errs := make(chan error, *numNodes)
start := time.Now()
deadline := start.Add(*duration)
var count int64
counts := make([]int64, *numNodes)
clients := make([]struct {
sync.RWMutex
db *client.DB
stopper *stop.Stopper
}, *numNodes)
initClient := func(i int) {
db, dbStopper := makeDBClient(t, l, i)
if clients[i].stopper != nil {
clients[i].stopper.Stop()
}
clients[i].db, clients[i].stopper = db, dbStopper
}
for i := 0; i < *numNodes; i++ {
initClient(i)
go func(i int) {
r, _ := randutil.NewPseudoRand()
value := randutil.RandBytes(r, 8192)
for time.Now().Before(deadline) {
clients[i].RLock()
k := atomic.AddInt64(&count, 1)
atomic.AddInt64(&counts[i], 1)
v := value[:r.Intn(len(value))]
if err := clients[i].db.Put(fmt.Sprintf("%08d", k), v); err != nil {
// These originate from DistSender when, for example, the
// leader is down. With more realistic retry options, we
// should probably not see them.
if _, ok := err.(*roachpb.SendError); ok {
log.Warning(err)
} else {
errs <- err
clients[i].RUnlock()
return
}
}
clients[i].RUnlock()
}
errs <- nil
}(i)
}
teardown := make(chan struct{})
defer func() {
<-teardown
for i := range clients {
clients[i].stopper.Stop()
clients[i].stopper = nil
}
}()
// Chaos monkey.
go func() {
defer close(teardown)
rnd, seed := randutil.NewPseudoRand()
log.Warningf("monkey starts (seed %d)", seed)
for round := 1; time.Now().Before(deadline); round++ {
select {
case <-stopper:
return
default:
}
nodes := rnd.Perm(*numNodes)[:rnd.Intn(*numNodes)+1]
log.Infof("round %d: restarting nodes %v", round, nodes)
for _, i := range nodes {
clients[i].Lock()
}
for _, i := range nodes {
log.Infof("restarting %v", i)
l.Nodes[i].Kill()
l.Nodes[i].Restart(5)
initClient(i)
clients[i].Unlock()
}
for cur := atomic.LoadInt64(&count); time.Now().Before(deadline) &&
atomic.LoadInt64(&count) == cur; time.Sleep(time.Second) {
l.Assert(t)
log.Warningf("monkey sleeping while cluster recovers...")
}
}
}()
for i := 0; i < *numNodes; {
select {
case <-teardown:
case <-stopper:
t.Fatal("interrupted")
case err := <-errs:
if err != nil {
t.Error(err)
}
i++
case <-time.After(1 * time.Second):
// Periodically print out progress so that we know the test is still
// running.
cur := make([]string, *numNodes)
for i := range cur {
cur[i] = fmt.Sprintf("%d", atomic.LoadInt64(&counts[i]))
}
log.Infof("%d (%s)", atomic.LoadInt64(&count), strings.Join(cur, ", "))
}
}
elapsed := time.Since(start)
log.Infof("%d %.1f/sec", count, float64(count)/elapsed.Seconds())
}