func (mvcc *MVCC) putInternal(key Key, timestamp hlc.Timestamp, value []byte, txnID string) error {
keyMeta := &keyMetadata{}
ok, err := GetI(mvcc.engine, key, keyMeta)
if err != nil {
return err
}
// In case the key metadata exists.
if ok {
// There is an uncommitted write intent and the current Put
// operation does not come from the same transaction.
// This should not happen since range should check the existing
// write intent before executing any Put action at MVCC level.
if len(keyMeta.TxnID) > 0 && (len(txnID) == 0 || keyMeta.TxnID != txnID) {
return &writeIntentError{TxnID: keyMeta.TxnID}
}
if keyMeta.Timestamp.Less(timestamp) ||
(timestamp.Equal(keyMeta.Timestamp) && txnID == keyMeta.TxnID) {
// Update key metadata.
PutI(mvcc.engine, key, &keyMetadata{TxnID: txnID, Timestamp: timestamp})
} else {
// In case we receive a Put request to update an old version,
// it must be an error since raft should handle any client
// retry from timeout.
return &writeTimestampTooOldError{Timestamp: keyMeta.Timestamp}
}
} else { // In case the key metadata does not exist yet.
// Create key metadata.
PutI(mvcc.engine, key, &keyMetadata{TxnID: txnID, Timestamp: timestamp})
}
// Save the value with the given version (Key + Timestamp).
return mvcc.engine.Put(mvccEncodeKey(key, timestamp), value)
}
// getInternal implements the actual logic of get function.
// The values of multiple versions for the given key should
// be organized as follows:
// ...
// keyA : keyMetatata of keyA
// keyA_Timestamp_n : value of version_n
// keyA_Timestamp_n-1 : value of version_n-1
// ...
// keyA_Timestamp_0 : value of version_0
// keyB : keyMetadata of keyB
// ...
func (mvcc *MVCC) getInternal(key Key, timestamp hlc.Timestamp, txnID string) ([]byte, hlc.Timestamp, string, error) {
keyMetadata := &keyMetadata{}
ok, err := GetI(mvcc.engine, key, keyMetadata)
if err != nil || !ok {
return nil, hlc.Timestamp{}, "", err
}
// If the read timestamp is greater than the latest one, we can just
// fetch the value without a scan.
if !timestamp.Less(keyMetadata.Timestamp) {
if len(keyMetadata.TxnID) > 0 && (len(txnID) == 0 || keyMetadata.TxnID != txnID) {
return nil, hlc.Timestamp{}, "", &writeIntentError{TxnID: keyMetadata.TxnID}
}
latestKey := mvccEncodeKey(key, keyMetadata.Timestamp)
val, err := mvcc.engine.Get(latestKey)
return val, keyMetadata.Timestamp, keyMetadata.TxnID, err
}
nextKey := mvccEncodeKey(key, timestamp)
// We use the PrefixEndKey(key) as the upper bound for scan.
// If there is no other version after nextKey, it won't return
// the value of the next key.
kvs, err := mvcc.engine.Scan(nextKey, PrefixEndKey(key), 1)
if len(kvs) > 0 {
_, ts := mvccDecodeKey(kvs[0].Key)
return kvs[0].Value, ts, "", err
}
return nil, hlc.Timestamp{}, "", err
}
// UpdateDeadlineMaybe sets the transactions deadline to the lower of the
// current one (if any) and the passed value.
func (txn *Txn) UpdateDeadlineMaybe(deadline hlc.Timestamp) bool {
if txn.deadline == nil || deadline.Less(*txn.deadline) {
txn.deadline = &deadline
return true
}
return false
}
func TestBatchBuilderStress(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
defer stopper.Stop()
e := NewInMem(roachpb.Attributes{}, 1<<20, stopper)
rng, _ := randutil.NewPseudoRand()
for i := 0; i < 1000; i++ {
count := 1 + rng.Intn(1000)
func() {
batch := e.NewBatch().(*rocksDBBatch)
defer batch.Close()
builder := &rocksDBBatchBuilder{}
for j := 0; j < count; j++ {
var ts hlc.Timestamp
if rng.Float32() <= 0.9 {
// Give 90% of keys timestamps.
ts.WallTime = rng.Int63()
if rng.Float32() <= 0.1 {
// Give 10% of timestamps a non-zero logical component.
ts.Logical = rng.Int31()
}
}
key := MVCCKey{
Key: []byte(fmt.Sprintf("%d", rng.Intn(10000))),
Timestamp: ts,
}
// Generate a random mixture of puts, deletes and merges.
switch rng.Intn(3) {
case 0:
if err := dbPut(batch.batch, key, []byte("value")); err != nil {
t.Fatal(err)
}
builder.Put(key, []byte("value"))
case 1:
if err := dbClear(batch.batch, key); err != nil {
t.Fatal(err)
}
builder.Clear(key)
case 2:
if err := dbMerge(batch.batch, key, appender("bar")); err != nil {
t.Fatal(err)
}
builder.Merge(key, appender("bar"))
}
}
batchRepr := batch.Repr()
builderRepr := builder.Finish()
if !bytes.Equal(batchRepr, builderRepr) {
t.Fatalf("expected [% x], but got [% x]", batchRepr, builderRepr)
}
}()
}
}
// UpdateObservedTimestamp stores a timestamp off a node's clock for future
// operations in the transaction. When multiple calls are made for a single
// nodeID, the lowest timestamp prevails.
func (t *Transaction) UpdateObservedTimestamp(nodeID NodeID, maxTS hlc.Timestamp) {
if t.ObservedTimestamps == nil {
t.ObservedTimestamps = make(map[NodeID]hlc.Timestamp)
}
if ts, ok := t.ObservedTimestamps[nodeID]; !ok || maxTS.Less(ts) {
t.ObservedTimestamps[nodeID] = maxTS
}
}
// isAsOf analyzes a select statement to bypass the logic in newPlan(),
// since that requires the transaction to be started already. If the returned
// timestamp is not nil, it is the timestamp to which a transaction should
// be set.
func isAsOf(planMaker *planner, stmt parser.Statement, max hlc.Timestamp) (*hlc.Timestamp, error) {
s, ok := stmt.(*parser.Select)
if !ok {
return nil, nil
}
sc, ok := s.Select.(*parser.SelectClause)
if !ok {
return nil, nil
}
if len(sc.From) != 1 {
return nil, nil
}
ate, ok := sc.From[0].(*parser.AliasedTableExpr)
if !ok {
return nil, nil
}
if ate.AsOf.Expr == nil {
return nil, nil
}
te, err := ate.AsOf.Expr.TypeCheck(nil, parser.TypeString)
if err != nil {
return nil, err
}
d, err := te.Eval(&planMaker.evalCtx)
if err != nil {
return nil, err
}
ds, ok := d.(*parser.DString)
if !ok {
return nil, fmt.Errorf("AS OF SYSTEM TIME expected string, got %s", ds.Type())
}
// Allow nanosecond precision because the timestamp is only used by the
// system and won't be returned to the user over pgwire.
dt, err := parser.ParseDTimestamp(string(*ds), planMaker.session.Location, time.Nanosecond)
if err != nil {
return nil, err
}
ts := hlc.Timestamp{
WallTime: dt.Time.UnixNano(),
}
if max.Less(ts) {
return nil, fmt.Errorf("cannot specify timestamp in the future")
}
return &ts, nil
}
func replicaGCShouldQueueImpl(
now, lastCheck, lastActivity hlc.Timestamp, isCandidate bool,
) (bool, float64) {
timeout := ReplicaGCQueueInactivityThreshold
var priority float64
if isCandidate {
// If the range is a candidate (which happens if its former replica set
// ignores it), let it expire much earlier.
timeout = ReplicaGCQueueCandidateTimeout
priority++
} else if now.Less(lastCheck.Add(ReplicaGCQueueInactivityThreshold.Nanoseconds(), 0)) {
// Return false immediately if the previous check was less than the
// check interval in the past. Note that we don't do this is the
// replica is in candidate state, in which case we want to be more
// aggressive - a failed rebalance attempt could have checked this
// range, and candidate state suggests that a retry succeeded. See
// #7489.
return false, 0
}
shouldQ := lastActivity.Add(timeout.Nanoseconds(), 0).Less(now)
if !shouldQ {
return false, 0
}
return shouldQ, priority
}
// selectEventTimestamp selects a timestamp for this log message. If the
// transaction this event is being written in has a non-zero timestamp, then that
// timestamp should be used; otherwise, the store's physical clock is used.
// This helps with testing; in normal usage, the logging of an event will never
// be the first action in the transaction, and thus the transaction will have an
// assigned database timestamp. However, in the case of our tests log events
// *are* the first action in a transaction, and we must elect to use the store's
// physical time instead.
func (ev EventLogger) selectEventTimestamp(input hlc.Timestamp) time.Time {
if input == hlc.ZeroTimestamp {
return ev.LeaseManager.clock.PhysicalTime()
}
return input.GoTime()
}
// selectEventTimestamp selects a timestamp for this log message. If the
// transaction this event is being written in has a non-zero timestamp, then that
// timestamp should be used; otherwise, the store's physical clock is used.
// This helps with testing; in normal usage, the logging of an event will never
// be the first action in the transaction, and thus the transaction will have an
// assigned database timestamp. However, in the case of our tests log events
// *are* the first action in a transaction, and we must elect to use the store's
// physical time instead.
func selectEventTimestamp(s *Store, input hlc.Timestamp) time.Time {
if input == hlc.ZeroTimestamp {
return s.Clock().PhysicalTime()
}
return input.GoTime()
}
// InitOrJoinRequest executes a RequestLease command asynchronously and returns a
// channel on which the result will be posted. If there's already a request in
// progress, we join in waiting for the results of that request.
// It is an error to call InitOrJoinRequest() while a request is in progress
// naming another replica as lease holder.
//
// replica is used to schedule and execute async work (proposing a RequestLease
// command). replica.mu is locked when delivering results, so calls from the
// replica happen either before or after a result for a pending request has
// happened.
//
// transfer needs to be set if the request represents a lease transfer (as
// opposed to an extension, or acquiring the lease when none is held).
//
// Note: Once this function gets a context to be used for cancellation, instead
// of replica.store.Stopper().ShouldQuiesce(), care will be needed for cancelling
// the Raft command, similar to replica.addWriteCmd.
func (p *pendingLeaseRequest) InitOrJoinRequest(
replica *Replica,
nextLeaseHolder roachpb.ReplicaDescriptor,
timestamp hlc.Timestamp,
startKey roachpb.Key,
transfer bool,
) <-chan *roachpb.Error {
if nextLease := p.RequestPending(); nextLease != nil {
if nextLease.Replica.ReplicaID == nextLeaseHolder.ReplicaID {
// Join a pending request asking for the same replica to become lease
// holder.
return p.JoinRequest()
}
llChan := make(chan *roachpb.Error, 1)
// We can't join the request in progress.
llChan <- roachpb.NewErrorf("request for different replica in progress "+
"(requesting: %+v, in progress: %+v)",
nextLeaseHolder.ReplicaID, nextLease.Replica.ReplicaID)
return llChan
}
llChan := make(chan *roachpb.Error, 1)
// No request in progress. Let's propose a Lease command asynchronously.
// TODO(tschottdorf): get duration from configuration, either as a
// config flag or, later, dynamically adjusted.
startStasis := timestamp.Add(int64(replica.store.ctx.rangeLeaseActiveDuration), 0)
expiration := startStasis.Add(int64(replica.store.Clock().MaxOffset()), 0)
reqSpan := roachpb.Span{
Key: startKey,
}
var leaseReq roachpb.Request
reqLease := roachpb.Lease{
Start: timestamp,
StartStasis: startStasis,
Expiration: expiration,
Replica: nextLeaseHolder,
}
if transfer {
leaseReq = &roachpb.TransferLeaseRequest{
Span: reqSpan,
Lease: reqLease,
}
} else {
leaseReq = &roachpb.RequestLeaseRequest{
Span: reqSpan,
Lease: reqLease,
}
}
if replica.store.Stopper().RunAsyncTask(func() {
// Propose a RequestLease command and wait for it to apply.
var execPErr *roachpb.Error
ba := roachpb.BatchRequest{}
ba.Timestamp = replica.store.Clock().Now()
ba.RangeID = replica.RangeID
ba.Add(leaseReq)
// Send lease request directly to raft in order to skip unnecessary
// checks from normal request machinery, (e.g. the command queue).
// Note that the command itself isn't traced, but usually the caller
// waiting for the result has an active Trace.
ch, _, err := replica.proposeRaftCommand(
replica.context(context.Background()), ba)
if err != nil {
execPErr = roachpb.NewError(err)
} else {
// If the command was committed, wait for the range to apply it.
select {
case c := <-ch:
if c.Err != nil {
if log.V(1) {
log.Infof("failed to acquire lease for replica %s: %s", replica.store, c.Err)
}
execPErr = c.Err
}
case <-replica.store.Stopper().ShouldQuiesce():
execPErr = roachpb.NewError(
replica.newNotLeaseHolderError(nil, replica.store.StoreID(), replica.Desc()))
}
}
// Send result of lease to all waiter channels.
replica.mu.Lock()
defer replica.mu.Unlock()
for i, llChan := range p.llChans {
// Don't send the same pErr object twice; this can lead to races. We could
// clone every time but it's more efficient to send pErr itself to one of
// the channels (the last one; if we send it earlier the race can still
// happen).
if i == len(p.llChans)-1 {
llChan <- execPErr
} else {
llChan <- protoutil.Clone(execPErr).(*roachpb.Error) // works with `nil`
}
}
p.llChans = p.llChans[:0]
p.nextLease = roachpb.Lease{}
}) != nil {
// We failed to start the asynchronous task. Send a blank NotLeaseHolderError
// back to indicate that we have no idea who the range lease holder might
// be; we've withdrawn from active duty.
llChan <- roachpb.NewError(
replica.newNotLeaseHolderError(nil, replica.store.StoreID(), replica.mu.state.Desc))
return llChan
}
p.llChans = append(p.llChans, llChan)
p.nextLease = reqLease
return llChan
}
// Covers returns true if the given timestamp can be served by the Lease.
// This is the case if the timestamp precedes the Lease's stasis period.
// Note that the fact that a lease convers a timestamp is not enough for the
// holder of the lease to be able to serve a read with that timestamp;
// pendingLeaderLeaseRequest.TransferInProgress() should also be consulted to
// account for possible lease transfers.
func (l Lease) Covers(timestamp hlc.Timestamp) bool {
return timestamp.Less(l.StartStasis)
}
// add the specified timestamp to the cache as covering the range of
// keys from start to end. If end is nil, the range covers the start
// key only. txnID is nil for no transaction. readTSCache specifies
// whether the command adding this timestamp should update the read
// timestamp; false to update the write timestamp cache.
func (tc *timestampCache) add(
start, end roachpb.Key,
timestamp hlc.Timestamp,
txnID *uuid.UUID,
readTSCache bool,
) {
// This gives us a memory-efficient end key if end is empty.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
tc.latest.Forward(timestamp)
// Only add to the cache if the timestamp is more recent than the
// low water mark.
if tc.lowWater.Less(timestamp) {
tcache := tc.wCache
if readTSCache {
tcache = tc.rCache
}
addRange := func(r interval.Range) {
value := cacheValue{timestamp: timestamp, txnID: txnID}
key := tcache.MakeKey(r.Start, r.End)
entry := makeCacheEntry(key, value)
tcache.AddEntry(entry)
}
r := interval.Range{
Start: interval.Comparable(start),
End: interval.Comparable(end),
}
// Check existing, overlapping entries and truncate/split/remove if
// superseded and in the past. If existing entries are in the future,
// subtract from the range/ranges that need to be added to cache.
for _, entry := range tcache.GetOverlaps(r.Start, r.End) {
cv := entry.Value.(*cacheValue)
key := entry.Key.(*cache.IntervalKey)
sCmp := r.Start.Compare(key.Start)
eCmp := r.End.Compare(key.End)
if !timestamp.Less(cv.timestamp) {
// The existing interval has a timestamp less than or equal to the new interval.
// Compare interval ranges to determine how to modify existing interval.
switch {
case sCmp == 0 && eCmp == 0:
// New and old are equal; replace old with new and avoid the need to insert new.
//
// New: ------------
// Old: ------------
//
// New: ------------
*cv = cacheValue{timestamp: timestamp, txnID: txnID}
tcache.MoveToEnd(entry)
return
case sCmp <= 0 && eCmp >= 0:
// New contains or is equal to old; delete old.
//
// New: ------------ ------------ ------------
// Old: -------- or ---------- or ----------
//
// Old:
tcache.DelEntry(entry)
case sCmp > 0 && eCmp < 0:
// Old contains new; split up old into two.
//
// New: ----
// Old: ------------
//
// Old: ---- ----
oldEnd := key.End
key.End = r.Start
key := tcache.MakeKey(r.End, oldEnd)
newEntry := makeCacheEntry(key, *cv)
tcache.AddEntryAfter(newEntry, entry)
case eCmp >= 0:
// Left partial overlap; truncate old end.
//
// New: -------- --------
// Old: -------- or ------------
//
// Old: ---- ----
key.End = r.Start
case sCmp <= 0:
// Right partial overlap; truncate old start.
//
// New: -------- --------
// Old: -------- or ------------
//
// Old: ---- ----
key.Start = r.End
default:
panic(fmt.Sprintf("no overlap between %v and %v", key.Range, r))
}
} else {
// The existing interval has a timestamp greater than the new interval.
// Compare interval ranges to determine how to modify new interval before
// adding it to the timestamp cache.
switch {
case sCmp >= 0 && eCmp <= 0:
// Old contains or is equal to new; no need to add.
//
// Old: ----------- ----------- ----------- -----------
// New: ----- or ----------- or -------- or --------
//
// New:
return
case sCmp < 0 && eCmp > 0:
// New contains old; split up old into two. We can add the left piece
// immediately because it is guaranteed to be before the rest of the
// overlaps.
//
// Old: ------
// New: ------------
//
// New: --- ---
lr := interval.Range{Start: r.Start, End: key.Start}
addRange(lr)
r.Start = key.End
case eCmp > 0:
// Left partial overlap; truncate new start.
//
// Old: -------- --------
// New: -------- or ------------
//
// New: ---- ----
r.Start = key.End
case sCmp < 0:
// Right partial overlap; truncate new end.
//
// Old: -------- --------
// New: -------- or ------------
//
// New: ---- ----
r.End = key.Start
default:
panic(fmt.Sprintf("no overlap between %v and %v", key.Range, r))
}
}
}
addRange(r)
}
}
// TestTxnCoordSenderHeartbeat verifies periodic heartbeat of the
// transaction record.
func TestTxnCoordSenderHeartbeat(t *testing.T) {
defer leaktest.AfterTest(t)()
s, sender := createTestDB(t)
defer s.Stop()
defer teardownHeartbeats(sender)
// Set heartbeat interval to 1ms for testing.
sender.heartbeatInterval = 1 * time.Millisecond
initialTxn := client.NewTxn(context.Background(), *s.DB)
if err := initialTxn.Put(roachpb.Key("a"), []byte("value")); err != nil {
t.Fatal(err)
}
// Verify 3 heartbeats.
var heartbeatTS hlc.Timestamp
for i := 0; i < 3; i++ {
util.SucceedsSoon(t, func() error {
txn, pErr := getTxn(sender, &initialTxn.Proto)
if pErr != nil {
t.Fatal(pErr)
}
// Advance clock by 1ns.
// Locking the TxnCoordSender to prevent a data race.
sender.Lock()
s.Manual.Increment(1)
sender.Unlock()
if txn.LastHeartbeat != nil && heartbeatTS.Less(*txn.LastHeartbeat) {
heartbeatTS = *txn.LastHeartbeat
return nil
}
return errors.Errorf("expected heartbeat")
})
}
// Sneakily send an ABORT right to DistSender (bypassing TxnCoordSender).
{
var ba roachpb.BatchRequest
ba.Add(&roachpb.EndTransactionRequest{
Commit: false,
Span: roachpb.Span{Key: initialTxn.Proto.Key},
})
ba.Txn = &initialTxn.Proto
if _, pErr := sender.wrapped.Send(context.Background(), ba); pErr != nil {
t.Fatal(pErr)
}
}
util.SucceedsSoon(t, func() error {
sender.Lock()
defer sender.Unlock()
if txnMeta, ok := sender.txns[*initialTxn.Proto.ID]; !ok {
t.Fatal("transaction unregistered prematurely")
} else if txnMeta.txn.Status != roachpb.ABORTED {
return fmt.Errorf("transaction is not aborted")
}
return nil
})
// Trying to do something else should give us a TransactionAbortedError.
_, err := initialTxn.Get("a")
assertTransactionAbortedError(t, err)
}