// mustGetInteger decodes an int64 value from the bytes field of the receiver
// and panics if the bytes field is not 0 or 8 bytes in length.
func mustGetInteger(v *proto.Value) int64 {
i, err := v.GetInteger()
if err != nil {
panic(err)
}
return i
}
// ConditionalPut sets the value for a specified key only if
// the expected value matches. If not, the return value contains
// the actual value.
func (mvcc *MVCC) ConditionalPut(key Key, timestamp proto.Timestamp, value proto.Value, expValue *proto.Value, txn *proto.Transaction) (*proto.Value, error) {
// Handle check for non-existence of key. In order to detect
// the potential write intent by another concurrent transaction
// with a newer timestamp, we need to use the max timestamp
// while reading.
existVal, err := mvcc.Get(key, proto.MaxTimestamp, txn)
if err != nil {
return nil, err
}
if expValue == nil && existVal != nil {
return existVal, util.Errorf("key %q already exists", key)
} else if expValue != nil {
// Handle check for existence when there is no key.
if existVal == nil {
return nil, util.Errorf("key %q does not exist", key)
} else if expValue.Bytes != nil && !bytes.Equal(expValue.Bytes, existVal.Bytes) {
return existVal, util.Errorf("key %q does not match existing", key)
} else if expValue.Integer != nil && (existVal.Integer == nil || expValue.GetInteger() != existVal.GetInteger()) {
return existVal, util.Errorf("key %q does not match existing", key)
}
}
return nil, mvcc.Put(key, timestamp, value, txn)
}
// putInternal writes the specified value to key.
func (kv *KV) putInternal(key proto.Key, value proto.Value) error {
value.InitChecksum(key)
return kv.Call(proto.Put, &proto.PutRequest{
RequestHeader: proto.RequestHeader{Key: key},
Value: value,
}, &proto.PutResponse{})
}
// PutProto sets the given key to the protobuf-serialized byte string
// of msg and the provided timestamp.
func (mvcc *MVCC) PutProto(key Key, timestamp proto.Timestamp, txn *proto.Transaction, msg gogoproto.Message) error {
data, err := gogoproto.Marshal(msg)
if err != nil {
return err
}
value := proto.Value{Bytes: data}
value.InitChecksum(key)
return mvcc.Put(key, timestamp, value, txn)
}
// mustGetInt decodes an int64 value from the bytes field of the receiver
// and panics if the bytes field is not 0 or 8 bytes in length.
func mustGetInt(v *proto.Value) int64 {
if v == nil {
return 0
}
i, err := v.GetInt()
if err != nil {
panic(err)
}
return i
}
// Indirectly this tests that the transaction remembers the NodeID of the node
// being read from correctly, at least in this simple case. Not remembering the
// node would lead to thousands of transaction restarts and almost certainly a
// test timeout.
func TestUncertaintyRestarts(t *testing.T) {
{
db, eng, clock, mClock, _, transport, err := createTestDB()
if err != nil {
t.Fatal(err)
}
defer transport.Close()
// Set a large offset so that a busy restart-loop
// really shows. Also makes sure that the values
// we write in the future below don't actually
// wind up in the past.
offset := 4000 * time.Millisecond
clock.SetMaxOffset(offset)
key := proto.Key("key")
value := proto.Value{
Bytes: nil, // Set for each Put
}
// With the correct restart behaviour, we see only one restart
// and the value read is the very first one (as nothing else
// has been written)
wantedBytes := []byte("value-0")
txnOpts := &client.TransactionOptions{
Name: "uncertainty",
}
gr := &proto.GetResponse{}
i := -1
tErr := db.RunTransaction(txnOpts, func(txn *client.KV) error {
i++
mClock.Increment(1)
futureTS := clock.Now()
futureTS.WallTime++
value.Bytes = []byte(fmt.Sprintf("value-%d", i))
err = engine.MVCCPut(eng, nil, key, futureTS, value, nil)
if err != nil {
t.Fatal(err)
}
gr.Reset()
if err := txn.Call(proto.Get, proto.GetArgs(key), gr); err != nil {
return err
}
if gr.Value == nil || !bytes.Equal(gr.Value.Bytes, wantedBytes) {
t.Fatalf("%d: read wrong value: %v, wanted %q", i,
gr.Value, wantedBytes)
}
return nil
})
if i != 1 {
t.Errorf("txn restarted %d times, expected only one restart", i)
}
if tErr != nil {
t.Fatal(tErr)
}
}
}
// putInternal writes the specified value to key.
func putInternal(db DB, key engine.Key, value proto.Value, timestamp proto.Timestamp) error {
value.InitChecksum(key)
pr := <-db.Put(&proto.PutRequest{
RequestHeader: proto.RequestHeader{
Key: key,
User: UserRoot,
Timestamp: timestamp,
},
Value: value,
})
return pr.GoError()
}
// PreparePutProto sets the given key to the protobuf-serialized byte
// string of msg. The resulting Put call is buffered and will not be
// sent until a subsequent call to Flush. Returns marshalling errors
// if encountered.
func (kv *KV) PreparePutProto(key proto.Key, msg gogoproto.Message) error {
data, err := gogoproto.Marshal(msg)
if err != nil {
return err
}
value := proto.Value{Bytes: data}
value.InitChecksum(key)
kv.Prepare(proto.Put, &proto.PutRequest{
RequestHeader: proto.RequestHeader{Key: key},
Value: value,
}, &proto.PutResponse{})
return nil
}
// 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>".
func setupMVCCScanData(numVersions, numKeys int, b *testing.B) (*RocksDB, *stop.Stopper) {
loc := fmt.Sprintf("mvcc_scan_%d_%d", numVersions, numKeys)
exists := true
if _, err := os.Stat(loc); os.IsNotExist(err) {
exists = false
}
log.Infof("creating mvcc data: %s", loc)
const cacheSize = 8 << 30 // 8 GB
stopper := stop.NewStopper()
rocksdb := NewRocksDB(proto.Attributes{Attrs: []string{"ssd"}}, 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
}
rng, _ := randutil.NewPseudoRand()
keys := make([]proto.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] = proto.Key(encoding.EncodeUvarint([]byte("key-"), uint64(i)))
nvs[i] = int(rand.Int31n(int32(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 := proto.Value{Bytes: randutil.RandBytes(rng, 1024)}
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
}
// Indirectly this tests that the transaction remembers the NodeID of the node
// being read from correctly, at least in this simple case. Not remembering the
// node would lead to thousands of transaction restarts and almost certainly a
// test timeout.
func TestUncertaintyRestarts(t *testing.T) {
defer leaktest.AfterTest(t)
s := createTestDB(t)
defer s.Stop()
// Set a large offset so that a busy restart-loop
// really shows. Also makes sure that the values
// we write in the future below don't actually
// wind up in the past.
offset := 4000 * time.Millisecond
s.Clock.SetMaxOffset(offset)
key := proto.Key("key")
value := proto.Value{
Bytes: nil, // Set for each Put
}
// With the correct restart behaviour, we see only one restart
// and the value read is the very first one (as nothing else
// has been written)
wantedBytes := []byte("value-0")
i := -1
tErr := s.DB.Txn(func(txn *client.Txn) error {
i++
s.Manual.Increment(1)
futureTS := s.Clock.Now()
futureTS.WallTime++
value.Bytes = []byte(fmt.Sprintf("value-%d", i))
if err := engine.MVCCPut(s.Eng, nil, key, futureTS, value, nil); err != nil {
t.Fatal(err)
}
gr, err := txn.Get(key)
if err != nil {
return err
}
if !gr.Exists() || !bytes.Equal(gr.ValueBytes(), wantedBytes) {
t.Fatalf("%d: read wrong value: %v, wanted %q", i, gr.Value, wantedBytes)
}
return nil
})
if i != 1 {
t.Errorf("txn restarted %d times, expected only one restart", i)
}
if tErr != nil {
t.Fatal(tErr)
}
}
// GetInitialSystemValues returns a list of key/value pairs.
// They are written at cluster bootstrap time (see storage/node.go:BootstrapCLuster).
func GetInitialSystemValues() []proto.KeyValue {
systemData := []struct {
parentID ID
desc descriptorProto
}{
{keys.RootNamespaceID, &SystemDB},
{SystemDB.ID, &NamespaceTable},
{SystemDB.ID, &DescriptorTable},
{SystemDB.ID, &UsersTable},
{SystemDB.ID, &ZonesTable},
}
// Initial kv pairs:
// - ID generator
// - 2 per table/database
numEntries := 1 + len(systemData)*2
ret := make([]proto.KeyValue, numEntries, numEntries)
i := 0
// Descriptor ID generator.
value := proto.Value{}
value.SetInt(int64(keys.MaxReservedDescID + 1))
ret[i] = proto.KeyValue{
Key: keys.DescIDGenerator,
Value: value,
}
i++
// System database and tables.
for _, d := range systemData {
value = proto.Value{}
value.SetInt(int64(d.desc.GetID()))
ret[i] = proto.KeyValue{
Key: MakeNameMetadataKey(d.parentID, d.desc.GetName()),
Value: value,
}
i++
value = proto.Value{}
if err := value.SetProto(d.desc); err != nil {
log.Fatalf("could not marshal %v", d.desc)
}
ret[i] = proto.KeyValue{
Key: MakeDescMetadataKey(d.desc.GetID()),
Value: value,
}
i++
}
return ret
}
// marshalValue returns a proto.Value initialized from the source
// reflect.Value, returning an error if the types are not compatible.
func marshalValue(v interface{}) (proto.Value, error) {
var r proto.Value
if v == nil {
return r, nil
}
switch t := v.(type) {
case nil:
return r, nil
case string:
r.SetBytes([]byte(t))
return r, nil
case []byte:
r.SetBytes(t)
return r, nil
case proto.Key:
r.SetBytes([]byte(t))
return r, nil
case time.Time:
err := r.SetTime(t)
return r, err
case gogoproto.Message:
err := r.SetProto(t)
return r, err
}
switch v := reflect.ValueOf(v); v.Kind() {
case reflect.Bool:
i := int64(0)
if v.Bool() {
i = 1
}
r.SetInt(i)
return r, nil
case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
r.SetInt(v.Int())
return r, nil
case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
r.SetInt(int64(v.Uint()))
return r, nil
case reflect.Float32, reflect.Float64:
r.SetFloat(v.Float())
return r, nil
case reflect.String:
r.SetBytes([]byte(v.String()))
return r, nil
}
return r, fmt.Errorf("unable to marshal value: %s", v)
}
// marshalValue returns a proto.Value initialized from the source
// reflect.Value, returning an error if the types are not compatible.
func marshalValue(v reflect.Value) (proto.Value, error) {
var r proto.Value
if !v.IsValid() {
return r, nil
}
switch t := v.Interface().(type) {
case nil:
return r, nil
case string:
r.Bytes = []byte(t)
return r, nil
case []byte:
r.Bytes = t
return r, nil
case proto.Key:
r.Bytes = []byte(t)
return r, nil
case gogoproto.Message:
var err error
r.Bytes, err = gogoproto.Marshal(t)
return r, err
case encoding.BinaryMarshaler:
var err error
r.Bytes, err = t.MarshalBinary()
return r, err
}
switch v.Kind() {
case reflect.Bool:
i := int64(0)
if v.Bool() {
i = 1
}
r.SetInteger(i)
return r, nil
case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
r.SetInteger(v.Int())
return r, nil
case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
r.SetInteger(int64(v.Uint()))
return r, nil
case reflect.Float32, reflect.Float64:
r.SetInteger(int64(math.Float64bits(v.Float())))
return r, nil
case reflect.String:
r.Bytes = []byte(v.String())
return r, nil
}
return r, fmt.Errorf("unable to marshal value: %s", v)
}
// unmarshalValue sets the destination reflect.Value contents from the source
// proto.Value, returning an error if the types are not compatible.
func unmarshalValue(src *proto.Value, dest reflect.Value) error {
if src == nil {
dest.Set(reflect.Zero(dest.Type()))
return nil
}
switch d := dest.Addr().Interface().(type) {
case *string:
if src.Bytes != nil {
*d = string(src.Bytes)
} else {
*d = ""
}
return nil
case *[]byte:
if src.Bytes != nil {
*d = src.Bytes
} else {
*d = nil
}
return nil
case *gogoproto.Message:
panic("TODO(pmattis): unimplemented")
case *encoding.BinaryUnmarshaler:
return (*d).UnmarshalBinary(src.Bytes)
}
switch dest.Kind() {
case reflect.Bool:
i, err := src.GetInteger()
if err != nil {
return err
}
dest.SetBool(i != 0)
return nil
case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
i, err := src.GetInteger()
if err != nil {
return err
}
dest.SetInt(i)
return nil
case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
i, err := src.GetInteger()
if err != nil {
return err
}
dest.SetUint(uint64(i))
return nil
case reflect.Float32, reflect.Float64:
i, err := src.GetInteger()
if err != nil {
return err
}
dest.SetFloat(math.Float64frombits(uint64(i)))
return nil
case reflect.String:
if src == nil || src.Bytes == nil {
dest.SetString("")
return nil
}
dest.SetString(string(src.Bytes))
return nil
}
return fmt.Errorf("unable to unmarshal value: %s", dest.Type())
}
// unmarshalColumnValue decodes the value from a key-value pair using the type
// expected by the column. An error is returned if the value's type does not
// match the column's type.
func unmarshalColumnValue(kind ColumnType_Kind, value *proto.Value) (parser.Datum, error) {
if value == nil {
return parser.DNull, nil
}
switch kind {
case ColumnType_BOOL:
v, err := value.GetInt()
if err != nil {
return nil, err
}
return parser.DBool(v != 0), nil
case ColumnType_INT:
v, err := value.GetInt()
if err != nil {
return nil, err
}
return parser.DInt(v), nil
case ColumnType_FLOAT:
v, err := value.GetFloat()
if err != nil {
return nil, err
}
return parser.DFloat(v), nil
case ColumnType_STRING:
v, err := value.GetBytesChecked()
if err != nil {
return nil, err
}
return parser.DString(v), nil
case ColumnType_BYTES:
v, err := value.GetBytesChecked()
if err != nil {
return nil, err
}
return parser.DBytes(v), nil
case ColumnType_DATE:
v, err := value.GetTime()
if err != nil {
return nil, err
}
return parser.DDate{Time: v}, nil
case ColumnType_TIMESTAMP:
v, err := value.GetTime()
if err != nil {
return nil, err
}
return parser.DTimestamp{Time: v}, nil
case ColumnType_INTERVAL:
v, err := value.GetInt()
if err != nil {
return nil, err
}
return parser.DInterval{Duration: time.Duration(v)}, nil
default:
return nil, util.Errorf("unsupported column type: %s", kind)
}
}
// marshalValue returns a proto.Value initialized from the source
// interface{}, returning an error if the types are not compatible.
func marshalValue(v interface{}) (proto.Value, error) {
var r proto.Value
// Handle a few common types via a type switch.
switch t := v.(type) {
case nil:
return r, nil
case bool:
i := int64(0)
if t {
i = 1
}
r.SetInt(i)
return r, nil
case string:
r.SetBytes([]byte(t))
return r, nil
case []byte:
r.SetBytes(t)
return r, nil
case proto.Key:
r.SetBytes([]byte(t))
return r, nil
case time.Time:
r.SetTime(t)
return r, nil
case gogoproto.Message:
err := r.SetProto(t)
return r, err
}
// Handle all of the Go primitive types besides struct and pointers. This
// switch also handles types based on a primitive type (e.g. "type MyInt
// int").
switch v := reflect.ValueOf(v); v.Kind() {
case reflect.Bool:
i := int64(0)
if v.Bool() {
i = 1
}
r.SetInt(i)
return r, nil
case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
r.SetInt(v.Int())
return r, nil
case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
r.SetInt(int64(v.Uint()))
return r, nil
case reflect.Float32, reflect.Float64:
r.SetFloat(v.Float())
return r, nil
case reflect.String:
r.SetBytes([]byte(v.String()))
return r, nil
}
return r, fmt.Errorf("unable to marshal value: %s", v)
}