// executeCmd interprets the given message as a *roachpb.BatchRequest and sends it
// via the local sender.
func (n *Node) executeCmd(argsI proto.Message) (proto.Message, error) {
ba := argsI.(*roachpb.BatchRequest)
var br *roachpb.BatchResponse
f := func() {
// TODO(tschottdorf) get a hold of the client's ID, add it to the
// context before dispatching, and create an ID for tracing the request.
sp := n.ctx.Tracer.StartSpan("node")
defer sp.Finish()
ctx, _ := opentracing.ContextWithSpan((*Node)(n).context(), sp)
tStart := time.Now()
var pErr *roachpb.Error
br, pErr = n.stores.Send(ctx, *ba)
if pErr != nil {
br = &roachpb.BatchResponse{}
sp.LogEvent(fmt.Sprintf("error: %T", pErr.GetDetail()))
}
if br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(n.stores, br))
}
n.feed.CallComplete(*ba, time.Now().Sub(tStart), pErr)
br.Error = pErr
}
if !n.stopper.RunTask(f) {
return nil, util.Errorf("node %d stopped", n.Descriptor.NodeID)
}
return br, nil
}
// If we hit an error and there is a pending transaction, rollback
// the transaction before returning. The client does not have to
// deal with cleaning up transaction state.
func makeResultFromError(planMaker *planner, pErr *roachpb.Error) Result {
if planMaker.txn != nil {
if _, ok := pErr.GetDetail().(*roachpb.SqlTransactionAbortedError); !ok {
planMaker.txn.Cleanup(pErr)
}
}
return Result{PErr: pErr}
}
// processWriteIntentError tries to push the conflicting
// transaction(s) responsible for the given WriteIntentError, and to
// resolve those intents if possible. Returns a new error to be used
// in place of the original.
//
// The returned error may be a copy of the original WriteIntentError,
// with or without the Resolved flag set, which governs the client's
// retry behavior (if the transaction is pushed, the Resolved flag is
// set to tell the client to retry immediately; otherwise it is false
// to cause the client to back off).
func (ir *intentResolver) processWriteIntentError(ctx context.Context,
wiPErr *roachpb.Error, args roachpb.Request, h roachpb.Header,
pushType roachpb.PushTxnType) *roachpb.Error {
wiErr, ok := wiPErr.GetDetail().(*roachpb.WriteIntentError)
if !ok {
return roachpb.NewErrorf("not a WriteIntentError: %v", wiPErr)
}
if log.V(6) {
log.Infof(ctx, "resolving write intent %s", wiErr)
}
method := args.Method()
readOnly := roachpb.IsReadOnly(args) // TODO(tschottdorf): pass as param
resolveIntents, pushErr := ir.maybePushTransactions(ctx, wiErr.Intents, h, pushType, false)
if resErr := ir.resolveIntents(ctx, resolveIntents,
false /* !wait */, pushType == roachpb.PUSH_ABORT /* poison */); resErr != nil {
// When resolving without waiting, errors should not
// usually be returned here, although there are some cases
// when they may be (especially when a test cluster is in
// the process of shutting down).
log.Warningf(ctx, "asynchronous resolveIntents failed: %s", resErr)
}
if pushErr != nil {
if log.V(1) {
log.Infof(ctx, "on %s: %s", method, pushErr)
}
if _, isExpected := pushErr.GetDetail().(*roachpb.TransactionPushError); !isExpected {
// If an unexpected error occurred, make sure it bubbles up to the
// client. Examples are timeouts and logic errors.
return pushErr
}
// For write/write conflicts within a transaction, propagate the
// push failure, not the original write intent error. The push
// failure will instruct the client to restart the transaction
// with a backoff.
if h.Txn != nil && h.Txn.ID != nil && !readOnly {
return pushErr
}
// For read/write conflicts, and non-transactional write/write
// conflicts, return the write intent error which engages
// backoff/retry (with !Resolved). We don't need to restart the
// txn, only resend the read with a backoff.
return wiPErr
}
// We pushed all transactions, so tell the client everything's
// resolved and it can retry immediately.
wiErr.Resolved = true
return wiPErr // references wiErr
}
func (c *v3Conn) sendPError(pErr *roachpb.Error) error {
var errCode string
if sqlErr, ok := pErr.GetDetail().(*roachpb.ErrorWithPGCode); ok {
errCode = sqlErr.ErrorCode
} else {
errCode = sql.CodeInternalError
}
return c.sendError(errCode, pErr.String())
}
// TestTxnCoordSenderEndTxn verifies that ending a transaction
// sends resolve write intent requests and removes the transaction
// from the txns map.
func TestTxnCoordSenderEndTxn(t *testing.T) {
defer leaktest.AfterTest(t)
s := createTestDB(t)
defer s.Stop()
// 4 cases: no deadline, past deadline, equal deadline, future deadline.
for i := 0; i < 4; i++ {
key := roachpb.Key("key: " + strconv.Itoa(i))
txn := client.NewTxn(*s.DB)
// Initialize the transaction
if pErr := txn.Put(key, []byte("value")); pErr != nil {
t.Fatal(pErr)
}
{
var pErr *roachpb.Error
switch i {
case 0:
// No deadline.
pErr = txn.Commit()
case 1:
// Past deadline.
pErr = txn.CommitBy(txn.Proto.Timestamp.Prev())
case 2:
// Equal deadline.
pErr = txn.CommitBy(txn.Proto.Timestamp)
case 3:
// Future deadline.
pErr = txn.CommitBy(txn.Proto.Timestamp.Next())
}
switch i {
case 0:
// No deadline.
if pErr != nil {
t.Error(pErr)
}
case 1:
// Past deadline.
if _, ok := pErr.GetDetail().(*roachpb.TransactionAbortedError); !ok {
t.Errorf("expected TransactionAbortedError but got %T: %s", pErr, pErr)
}
case 2:
// Equal deadline.
if pErr != nil {
t.Error(pErr)
}
case 3:
// Future deadline.
if pErr != nil {
t.Error(pErr)
}
}
}
verifyCleanup(key, s.Sender, s.Eng, t)
}
}
// executeCmd interprets the given message as a *roachpb.BatchRequest and sends it
// via the local sender.
func (n *Node) executeCmd(argsI proto.Message) (proto.Message, error) {
ba := argsI.(*roachpb.BatchRequest)
var br *roachpb.BatchResponse
opName := "node " + strconv.Itoa(int(n.Descriptor.NodeID)) // could save allocs here
fail := func(err error) {
br = &roachpb.BatchResponse{}
br.Error = roachpb.NewError(err)
}
f := func() {
sp, err := tracing.JoinOrNew(n.ctx.Tracer, ba.Trace, opName)
if err != nil {
fail(err)
return
}
// If this is a snowball span, it gets special treatment: It skips the
// regular tracing machinery, and we instead send the collected spans
// back with the response. This is more expensive, but then again,
// those are individual requests traced by users, so they can be.
if sp.BaggageItem(tracing.Snowball) != "" {
if sp, err = tracing.JoinOrNewSnowball(opName, ba.Trace, func(rawSpan basictracer.RawSpan) {
encSp, err := tracing.EncodeRawSpan(&rawSpan, nil)
if err != nil {
log.Warning(err)
}
br.CollectedSpans = append(br.CollectedSpans, encSp)
}); err != nil {
fail(err)
return
}
}
defer sp.Finish()
ctx := opentracing.ContextWithSpan((*Node)(n).context(), sp)
tStart := time.Now()
var pErr *roachpb.Error
br, pErr = n.stores.Send(ctx, *ba)
if pErr != nil {
br = &roachpb.BatchResponse{}
sp.LogEvent(fmt.Sprintf("error: %T", pErr.GetDetail()))
}
if br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(n.stores, br))
}
n.metrics.callComplete(time.Now().Sub(tStart), pErr)
br.Error = pErr
}
if !n.stopper.RunTask(f) {
return nil, util.Errorf("node %d stopped", n.Descriptor.NodeID)
}
return br, nil
}
// Responses with write-intent and not leader errors are retried on
// the server, and so are not recorded in the sequence cache in the
// hopes of retrying to a successful outcome.
func (sc *SequenceCache) shouldCacheError(pErr *roachpb.Error) bool {
switch pErr.GetDetail().(type) {
case *roachpb.WriteIntentError, *roachpb.NotLeaderError, *roachpb.RangeKeyMismatchError:
return false
}
return true
}
func convertBatchError(tableDesc *sqlbase.TableDescriptor, b client.Batch, origPErr *roachpb.Error) error {
if origPErr.Index == nil {
return origPErr.GoError()
}
index := origPErr.Index.Index
if index >= int32(len(b.Results)) {
panic(fmt.Sprintf("index %d outside of results: %+v", index, b.Results))
}
result := b.Results[index]
var alloc sqlbase.DatumAlloc
if _, ok := origPErr.GetDetail().(*roachpb.ConditionFailedError); ok {
for _, row := range result.Rows {
indexID, key, err := sqlbase.DecodeIndexKeyPrefix(tableDesc, row.Key)
if err != nil {
return err
}
index, err := tableDesc.FindIndexByID(indexID)
if err != nil {
return err
}
valTypes, err := sqlbase.MakeKeyVals(tableDesc, index.ColumnIDs)
if err != nil {
return err
}
dirs := make([]encoding.Direction, 0, len(index.ColumnIDs))
for _, dir := range index.ColumnDirections {
convertedDir, err := dir.ToEncodingDirection()
if err != nil {
return err
}
dirs = append(dirs, convertedDir)
}
vals := make([]parser.Datum, len(valTypes))
if _, err := sqlbase.DecodeKeyVals(&alloc, valTypes, vals, dirs, key); err != nil {
return err
}
return &errUniquenessConstraintViolation{index: index, vals: vals}
}
}
return origPErr.GoError()
}
// handlePerReplicaError returns true if the given error is likely to
// be unique to the replica that reported it, and retrying on other
// replicas is likely to produce different results. This method should
// be called only once for each error as it may have side effects such
// as updating caches.
func (ds *DistSender) handlePerReplicaError(rangeID roachpb.RangeID, pErr *roachpb.Error) bool {
switch tErr := pErr.GetDetail().(type) {
case *roachpb.RangeNotFoundError:
return true
case *roachpb.NodeUnavailableError:
return true
case *roachpb.NotLeaseHolderError:
if tErr.LeaseHolder != nil {
// If the replica we contacted knows the new lease holder, update the cache.
ds.updateLeaseHolderCache(rangeID, *tErr.LeaseHolder)
// TODO(bdarnell): Move the new lease holder to the head of the queue
// for the next retry.
}
return true
}
return false
}
// Send implements the batch.Sender interface. If the request is part of a
// transaction, the TxnCoordSender adds the transaction to a map of active
// transactions and begins heartbeating it. Every subsequent request for the
// same transaction updates the lastUpdate timestamp to prevent live
// transactions from being considered abandoned and garbage collected.
// Read/write mutating requests have their key or key range added to the
// transaction's interval tree of key ranges for eventual cleanup via resolved
// write intents; they're tagged to an outgoing EndTransaction request, with
// the receiving replica in charge of resolving them.
func (tc *TxnCoordSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
{
// Start new or pick up active trace and embed its trace metadata into
// header for use by RPC recipients. From here on, there's always an active
// Trace, though its overhead is small unless it's sampled.
sp := opentracing.SpanFromContext(ctx)
// TODO(radu): once contexts are plumbed correctly, we should use the Tracer
// from ctx.
tracer := tracing.TracerFromCtx(tc.ctx)
if sp == nil {
sp = tracer.StartSpan(opTxnCoordSender)
defer sp.Finish()
ctx = opentracing.ContextWithSpan(ctx, sp)
}
// TODO(tschottdorf): To get rid of the spurious alloc below we need to
// implement the carrier interface on ba.Header or make Span non-nullable,
// both of which force all of ba on the Heap. It's already there, so may
// not be a big deal, but ba should live on the stack. Also not easy to use
// a buffer pool here since anything that goes into the RPC layer could be
// used by goroutines we didn't wait for.
if ba.Header.Trace == nil {
ba.Header.Trace = &tracing.Span{}
} else {
// We didn't make this object but are about to mutate it, so we
// have to take a copy - the original might already have been
// passed to the RPC layer.
ba.Header.Trace = protoutil.Clone(ba.Header.Trace).(*tracing.Span)
}
if err := tracer.Inject(sp.Context(), basictracer.Delegator, ba.Trace); err != nil {
return nil, roachpb.NewError(err)
}
}
startNS := tc.clock.PhysicalNow()
if ba.Txn != nil {
// If this request is part of a transaction...
if err := tc.maybeBeginTxn(&ba); err != nil {
return nil, roachpb.NewError(err)
}
var et *roachpb.EndTransactionRequest
var hasET bool
{
var rArgs roachpb.Request
rArgs, hasET = ba.GetArg(roachpb.EndTransaction)
if hasET {
et = rArgs.(*roachpb.EndTransactionRequest)
if len(et.Key) != 0 {
return nil, roachpb.NewErrorf("EndTransaction must not have a Key set")
}
et.Key = ba.Txn.Key
if len(et.IntentSpans) > 0 {
// TODO(tschottdorf): it may be useful to allow this later.
// That would be part of a possible plan to allow txns which
// write on multiple coordinators.
return nil, roachpb.NewErrorf("client must not pass intents to EndTransaction")
}
}
}
if pErr := func() *roachpb.Error {
tc.Lock()
defer tc.Unlock()
if pErr := tc.maybeRejectClientLocked(ctx, *ba.Txn); pErr != nil {
return pErr
}
if !hasET {
return nil
}
// Everything below is carried out only when trying to commit.
// Populate et.IntentSpans, taking into account both any existing
// and new writes, and taking care to perform proper deduplication.
txnMeta := tc.txns[*ba.Txn.ID]
distinctSpans := true
if txnMeta != nil {
et.IntentSpans = txnMeta.keys
// Defensively set distinctSpans to false if we had any previous
// requests in this transaction. This effectively limits the distinct
// spans optimization to 1pc transactions.
distinctSpans = len(txnMeta.keys) == 0
}
ba.IntentSpanIterate(func(key, endKey roachpb.Key) {
et.IntentSpans = append(et.IntentSpans, roachpb.Span{
Key: key,
EndKey: endKey,
})
})
// TODO(peter): Populate DistinctSpans on all batches, not just batches
// which contain an EndTransactionRequest.
var distinct bool
// The request might already be used by an outgoing goroutine, so
// we can't safely mutate anything in-place (as MergeSpans does).
et.IntentSpans = append([]roachpb.Span(nil), et.IntentSpans...)
et.IntentSpans, distinct = roachpb.MergeSpans(et.IntentSpans)
ba.Header.DistinctSpans = distinct && distinctSpans
if len(et.IntentSpans) == 0 {
// If there aren't any intents, then there's factually no
// transaction to end. Read-only txns have all of their state
// in the client.
return roachpb.NewErrorf("cannot commit a read-only transaction")
}
if txnMeta != nil {
txnMeta.keys = et.IntentSpans
}
return nil
}(); pErr != nil {
return nil, pErr
}
if hasET && log.V(1) {
for _, intent := range et.IntentSpans {
log.Tracef(ctx, "intent: [%s,%s)", intent.Key, intent.EndKey)
}
}
}
// Send the command through wrapped sender, taking appropriate measures
// on error.
var br *roachpb.BatchResponse
{
var pErr *roachpb.Error
br, pErr = tc.wrapped.Send(ctx, ba)
if _, ok := pErr.GetDetail().(*roachpb.OpRequiresTxnError); ok {
// TODO(tschottdorf): needs to keep the trace.
br, pErr = tc.resendWithTxn(ba)
}
if pErr = tc.updateState(startNS, ctx, ba, br, pErr); pErr != nil {
log.Tracef(ctx, "error: %s", pErr)
return nil, pErr
}
}
if br.Txn == nil {
return br, nil
}
if _, ok := ba.GetArg(roachpb.EndTransaction); !ok {
return br, nil
}
// If the --linearizable flag is set, we want to make sure that
// all the clocks in the system are past the commit timestamp
// of the transaction. This is guaranteed if either
// - the commit timestamp is MaxOffset behind startNS
// - MaxOffset ns were spent in this function
// when returning to the client. Below we choose the option
// that involves less waiting, which is likely the first one
// unless a transaction commits with an odd timestamp.
if tsNS := br.Txn.Timestamp.WallTime; startNS > tsNS {
startNS = tsNS
}
sleepNS := tc.clock.MaxOffset() -
time.Duration(tc.clock.PhysicalNow()-startNS)
if tc.linearizable && sleepNS > 0 {
defer func() {
if log.V(1) {
log.Infof(ctx, "%v: waiting %s on EndTransaction for linearizability", br.Txn.ID.Short(), util.TruncateDuration(sleepNS, time.Millisecond))
}
time.Sleep(sleepNS)
}()
}
if br.Txn.Status != roachpb.PENDING {
tc.Lock()
tc.cleanupTxnLocked(ctx, *br.Txn)
tc.Unlock()
}
return br, nil
}
// TestPriorityRatchetOnAbortOrPush verifies that the priority of
// a transaction is ratcheted by successive aborts or pushes. In
// particular, we want to ensure ratcheted priorities when the txn
// discovers it's been aborted or pushed through a poisoned sequence
// cache. This happens when a concurrent writer aborts an intent or a
// concurrent reader pushes an intent.
func TestPriorityRatchetOnAbortOrPush(t *testing.T) {
defer leaktest.AfterTest(t)()
s := createTestDB(t)
defer s.Stop()
const pusheePri = 1
const pusherPri = 10 // pusher will win
pushByReading := func(key roachpb.Key) {
if pErr := s.DB.Txn(func(txn *client.Txn) *roachpb.Error {
txn.InternalSetPriority(pusherPri)
_, pErr := txn.Get(key)
return pErr
}); pErr != nil {
t.Fatal(pErr)
}
}
abortByWriting := func(key roachpb.Key) {
if pErr := s.DB.Txn(func(txn *client.Txn) *roachpb.Error {
txn.InternalSetPriority(pusherPri)
return txn.Put(key, "foo")
}); pErr != nil {
t.Fatal(pErr)
}
}
// Try all combinations of read/write and snapshot/serializable isolation.
for _, read := range []bool{true, false} {
for _, iso := range []roachpb.IsolationType{roachpb.SNAPSHOT, roachpb.SERIALIZABLE} {
var iteration int
if pErr := s.DB.Txn(func(txn *client.Txn) *roachpb.Error {
defer func() { iteration++ }()
key := roachpb.Key(fmt.Sprintf("read=%t, iso=%s", read, iso))
// Only set our priority on first try.
if iteration == 0 {
txn.InternalSetPriority(pusheePri)
}
if err := txn.SetIsolation(iso); err != nil {
t.Fatal(err)
}
// Write to lay down an intent (this will send the begin
// transaction which gets the updated priority).
if pErr := txn.Put(key, "bar"); pErr != nil {
return pErr
}
if iteration == 1 {
// Verify our priority has ratcheted to one less than the pusher's priority
if pri := txn.Proto.Priority; pri != pusherPri-1 {
t.Fatalf("%s: expected priority on retry to ratchet to %d; got %d", key, pusherPri-1, pri)
}
return nil
}
// Now simulate a concurrent reader or writer. Our txn will
// either be pushed or aborted. Then issue a read and verify
// that if we've been pushed, no error is returned and if we
// have been aborted, we get an aborted error.
var pErr *roachpb.Error
if read {
pushByReading(key)
_, pErr = txn.Get(key)
if pErr != nil {
t.Fatalf("%s: expected no error; got %s", key, pErr)
}
} else {
abortByWriting(key)
_, pErr = txn.Get(key)
if _, ok := pErr.GetDetail().(*roachpb.TransactionAbortedError); !ok {
t.Fatalf("%s: expected transaction aborted error; got %s", key, pErr)
}
}
return pErr
}); pErr != nil {
t.Fatal(pErr)
}
}
}
}
// updateState updates the transaction state in both the success and
// error cases, applying those updates to the corresponding txnMeta
// object when adequate. It also updates certain errors with the
// updated transaction for use by client restarts.
func (tc *TxnCoordSender) updateState(
startNS int64, ctx context.Context, ba roachpb.BatchRequest,
br *roachpb.BatchResponse, pErr *roachpb.Error) *roachpb.Error {
newTxn := &roachpb.Transaction{}
newTxn.Update(ba.Txn)
if pErr == nil {
newTxn.Update(br.Txn)
} else {
newTxn.Update(pErr.GetTxn())
}
switch t := pErr.GetDetail().(type) {
case *roachpb.TransactionStatusError:
// Likely already committed or more obscure errors such as epoch or
// timestamp regressions; consider txn dead.
defer tc.cleanupTxn(ctx, *pErr.GetTxn())
case *roachpb.OpRequiresTxnError:
panic("OpRequiresTxnError must not happen at this level")
case *roachpb.ReadWithinUncertaintyIntervalError:
// If the reader encountered a newer write within the uncertainty
// interval, we advance the txn's timestamp just past the last observed
// timestamp from the node.
restartTS, ok := newTxn.GetObservedTimestamp(pErr.OriginNode)
if !ok {
pErr = roachpb.NewError(util.Errorf("no observed timestamp for node %d found on uncertainty restart", pErr.OriginNode))
} else {
newTxn.Timestamp.Forward(restartTS)
newTxn.Restart(ba.UserPriority, newTxn.Priority, newTxn.Timestamp)
}
case *roachpb.TransactionAbortedError:
// Increase timestamp if applicable.
newTxn.Timestamp.Forward(pErr.GetTxn().Timestamp)
newTxn.Priority = pErr.GetTxn().Priority
// Clean up the freshly aborted transaction in defer(), avoiding a
// race with the state update below.
defer tc.cleanupTxn(ctx, *newTxn)
case *roachpb.TransactionPushError:
// Increase timestamp if applicable, ensuring that we're
// just ahead of the pushee.
newTxn.Timestamp.Forward(t.PusheeTxn.Timestamp)
newTxn.Restart(ba.UserPriority, t.PusheeTxn.Priority-1, newTxn.Timestamp)
case *roachpb.TransactionRetryError:
// Increase timestamp so on restart, we're ahead of any timestamp
// cache entries or newer versions which caused the restart.
newTxn.Restart(ba.UserPriority, pErr.GetTxn().Priority, newTxn.Timestamp)
case *roachpb.WriteTooOldError:
newTxn.Restart(ba.UserPriority, newTxn.Priority, t.ActualTimestamp)
case nil:
// Nothing to do here, avoid the default case.
default:
if pErr.GetTxn() != nil {
if pErr.CanRetry() {
panic("Retryable internal error must not happen at this level")
} else {
// Do not clean up the transaction here since the client might still
// want to continue the transaction. For example, a client might
// continue its transaction after receiving ConditionFailedError, which
// can come from a unique index violation.
}
}
}
if pErr != nil && pErr.GetTxn() != nil {
// Avoid changing existing errors because sometimes they escape into
// goroutines and then there are races. Fairly sure there isn't one
// here, but better safe than sorry.
pErrShallow := *pErr
pErrShallow.SetTxn(newTxn)
pErr = &pErrShallow
}
if newTxn.ID == nil {
return pErr
}
txnID := *newTxn.ID
tc.Lock()
defer tc.Unlock()
txnMeta := tc.txns[txnID]
// For successful transactional requests, keep the written intents and
// the updated transaction record to be sent along with the reply.
// The transaction metadata is created with the first writing operation.
// A tricky edge case is that of a transaction which "fails" on the
// first writing request, but actually manages to write some intents
// (for example, due to being multi-range). In this case, there will
// be an error, but the transaction will be marked as Writing and the
// coordinator must track the state, for the client's retry will be
// performed with a Writing transaction which the coordinator rejects
// unless it is tracking it (on top of it making sense to track it;
// after all, it **has** laid down intents and only the coordinator
// can augment a potential EndTransaction call). See #3303.
var intentGroup interval.RangeGroup
if txnMeta != nil {
intentGroup = txnMeta.keys
} else if pErr == nil || newTxn.Writing {
intentGroup = interval.NewRangeTree()
}
if intentGroup != nil {
// Adding the intents even on error reduces the likelihood of dangling
// intents blocking concurrent writers for extended periods of time.
// See #3346.
ba.IntentSpanIterate(func(key, endKey roachpb.Key) {
addKeyRange(intentGroup, key, endKey)
})
if txnMeta == nil && intentGroup.Len() > 0 {
if !newTxn.Writing {
panic("txn with intents marked as non-writing")
}
// If the transaction is already over, there's no point in
// launching a one-off coordinator which will shut down right
// away. If we ended up here with an error, we'll always start
// the coordinator - the transaction has laid down intents, so
// we expect it to be committed/aborted at some point in the
// future.
if _, isEnding := ba.GetArg(roachpb.EndTransaction); pErr != nil || !isEnding {
log.Trace(ctx, "coordinator spawns")
txnMeta = &txnMetadata{
txn: *newTxn,
keys: intentGroup,
firstUpdateNanos: startNS,
lastUpdateNanos: tc.clock.PhysicalNow(),
timeoutDuration: tc.clientTimeout,
txnEnd: make(chan struct{}),
}
tc.txns[txnID] = txnMeta
if !tc.stopper.RunAsyncTask(func() {
tc.heartbeatLoop(ctx, txnID)
}) {
// The system is already draining and we can't start the
// heartbeat. We refuse new transactions for now because
// they're likely not going to have all intents committed.
// In principle, we can relax this as needed though.
tc.unregisterTxnLocked(txnID)
return roachpb.NewError(&roachpb.NodeUnavailableError{})
}
} else {
// If this was a successful one phase commit, update stats
// directly as they won't otherwise be updated on heartbeat
// loop shutdown.
etArgs, ok := br.Responses[len(br.Responses)-1].GetInner().(*roachpb.EndTransactionResponse)
tc.updateStats(tc.clock.PhysicalNow()-startNS, 0, newTxn.Status, ok && etArgs.OnePhaseCommit)
}
}
}
// Update our record of this transaction, even on error.
if txnMeta != nil {
txnMeta.txn = *newTxn
if !txnMeta.txn.Writing {
panic("tracking a non-writing txn")
}
txnMeta.setLastUpdate(tc.clock.PhysicalNow())
}
if pErr == nil {
// For successful transactional requests, always send the updated txn
// record back.
br.Txn = newTxn
}
return pErr
}
// Send implements the batch.Sender interface. If the request is part of a
// transaction, the TxnCoordSender adds the transaction to a map of active
// transactions and begins heartbeating it. Every subsequent request for the
// same transaction updates the lastUpdate timestamp to prevent live
// transactions from being considered abandoned and garbage collected.
// Read/write mutating requests have their key or key range added to the
// transaction's interval tree of key ranges for eventual cleanup via resolved
// write intents; they're tagged to an outgoing EndTransaction request, with
// the receiving replica in charge of resolving them.
func (tc *TxnCoordSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
{
// Start new or pick up active trace and embed its trace metadata into
// header for use by RPC recipients. From here on, there's always an active
// Trace, though its overhead is small unless it's sampled.
sp := opentracing.SpanFromContext(ctx)
if sp == nil {
sp = tc.tracer.StartSpan(opTxnCoordSender)
defer sp.Finish()
ctx = opentracing.ContextWithSpan(ctx, sp)
}
// TODO(tschottdorf): To get rid of the spurious alloc below we need to
// implement the carrier interface on ba.Header or make Span non-nullable,
// both of which force all of ba on the Heap. It's already there, so may
// not be a big deal, but ba should live on the stack. Also not easy to use
// a buffer pool here since anything that goes into the RPC layer could be
// used by goroutines we didn't wait for.
if ba.Header.Trace == nil {
ba.Header.Trace = &tracing.Span{}
}
if err := tc.tracer.Inject(sp, basictracer.Delegator, ba.Trace); err != nil {
return nil, roachpb.NewError(err)
}
}
startNS := tc.clock.PhysicalNow()
if ba.Txn != nil {
// If this request is part of a transaction...
if err := tc.maybeBeginTxn(&ba); err != nil {
return nil, roachpb.NewError(err)
}
txnID := *ba.Txn.ID
// Verify that if this Transaction is not read-only, we have it on file.
// If not, refuse further operations - the transaction was aborted due
// to a timeout or the client must have issued a write on another
// coordinator previously.
if ba.Txn.Writing {
tc.Lock()
_, ok := tc.txns[txnID]
tc.Unlock()
if !ok {
pErr := roachpb.NewErrorf("writing transaction timed out, was aborted, " +
"or ran on multiple coordinators")
return nil, pErr
}
}
if rArgs, ok := ba.GetArg(roachpb.EndTransaction); ok {
et := rArgs.(*roachpb.EndTransactionRequest)
if len(et.Key) != 0 {
return nil, roachpb.NewErrorf("EndTransaction must not have a Key set")
}
et.Key = ba.Txn.Key
if len(et.IntentSpans) > 0 {
// TODO(tschottdorf): it may be useful to allow this later.
// That would be part of a possible plan to allow txns which
// write on multiple coordinators.
return nil, roachpb.NewErrorf("client must not pass intents to EndTransaction")
}
tc.Lock()
txnMeta, metaOK := tc.txns[txnID]
{
// Populate et.IntentSpans, taking into account both existing
// writes (if any) and new writes in this batch, and taking
// care to perform proper deduplication.
var keys interval.RangeGroup
if metaOK {
keys = txnMeta.keys
} else {
keys = interval.NewRangeTree()
}
ba.IntentSpanIterate(func(key, endKey roachpb.Key) {
addKeyRange(keys, key, endKey)
})
et.IntentSpans = collectIntentSpans(keys)
}
tc.Unlock()
if len(et.IntentSpans) > 0 {
// All good, proceed.
} else if !metaOK {
// If we don't have the transaction, then this must be a retry
// by the client. We can no longer reconstruct a correct
// request so we must fail.
//
// TODO(bdarnell): if we had a GetTransactionStatus API then
// we could lookup the transaction and return either nil or
// TransactionAbortedError instead of this ambivalent error.
return nil, roachpb.NewErrorf("transaction is already committed or aborted")
}
if len(et.IntentSpans) == 0 {
// If there aren't any intents, then there's factually no
// transaction to end. Read-only txns have all of their state in
// the client.
return nil, roachpb.NewErrorf("cannot commit a read-only transaction")
}
if log.V(1) {
for _, intent := range et.IntentSpans {
log.Trace(ctx, fmt.Sprintf("intent: [%s,%s)", intent.Key, intent.EndKey))
}
}
}
}
// Send the command through wrapped sender, taking appropriate measures
// on error.
var br *roachpb.BatchResponse
{
var pErr *roachpb.Error
br, pErr = tc.wrapped.Send(ctx, ba)
if _, ok := pErr.GetDetail().(*roachpb.OpRequiresTxnError); ok {
// TODO(tschottdorf): needs to keep the trace.
br, pErr = tc.resendWithTxn(ba)
}
if pErr = tc.updateState(startNS, ctx, ba, br, pErr); pErr != nil {
log.Trace(ctx, fmt.Sprintf("error: %s", pErr))
return nil, pErr
}
}
if br.Txn == nil {
return br, nil
}
if _, ok := ba.GetArg(roachpb.EndTransaction); !ok {
return br, nil
}
// If the --linearizable flag is set, we want to make sure that
// all the clocks in the system are past the commit timestamp
// of the transaction. This is guaranteed if either
// - the commit timestamp is MaxOffset behind startNS
// - MaxOffset ns were spent in this function
// when returning to the client. Below we choose the option
// that involves less waiting, which is likely the first one
// unless a transaction commits with an odd timestamp.
if tsNS := br.Txn.Timestamp.WallTime; startNS > tsNS {
startNS = tsNS
}
sleepNS := tc.clock.MaxOffset() -
time.Duration(tc.clock.PhysicalNow()-startNS)
if tc.linearizable && sleepNS > 0 {
defer func() {
if log.V(1) {
log.Infof("%v: waiting %s on EndTransaction for linearizability", br.Txn.ID.Short(), util.TruncateDuration(sleepNS, time.Millisecond))
}
time.Sleep(sleepNS)
}()
}
if br.Txn.Status != roachpb.PENDING {
tc.cleanupTxn(ctx, *br.Txn)
}
return br, nil
}
// TestClientRetryNonTxn verifies that non-transactional client will
// succeed despite write/write and read/write conflicts. In the case
// where the non-transactional put can push the txn, we expect the
// transaction's value to be written after all retries are complete.
func TestClientRetryNonTxn(t *testing.T) {
defer leaktest.AfterTest(t)
s := server.StartTestServer(t)
defer s.Stop()
s.SetRangeRetryOptions(retry.Options{
InitialBackoff: 1 * time.Millisecond,
MaxBackoff: 5 * time.Millisecond,
Multiplier: 2,
MaxRetries: 1,
})
testCases := []struct {
args roachpb.Request
isolation roachpb.IsolationType
canPush bool
expAttempts int
}{
// Write/write conflicts.
{&roachpb.PutRequest{}, roachpb.SNAPSHOT, true, 2},
{&roachpb.PutRequest{}, roachpb.SERIALIZABLE, true, 2},
{&roachpb.PutRequest{}, roachpb.SNAPSHOT, false, 1},
{&roachpb.PutRequest{}, roachpb.SERIALIZABLE, false, 1},
// Read/write conflicts.
{&roachpb.GetRequest{}, roachpb.SNAPSHOT, true, 1},
{&roachpb.GetRequest{}, roachpb.SERIALIZABLE, true, 2},
{&roachpb.GetRequest{}, roachpb.SNAPSHOT, false, 1},
{&roachpb.GetRequest{}, roachpb.SERIALIZABLE, false, 1},
}
// Lay down a write intent using a txn and attempt to write to same
// key. Try this twice--once with priorities which will allow the
// intent to be pushed and once with priorities which will not.
for i, test := range testCases {
key := roachpb.Key(fmt.Sprintf("key-%d", i))
var txnPri int32 = 1
var clientPri roachpb.UserPriority = 1
if test.canPush {
clientPri = 2
} else {
txnPri = 2
}
db, sender := createTestNotifyClient(s.Stopper(), s.ServingAddr(), -clientPri)
// doneCall signals when the non-txn read or write has completed.
doneCall := make(chan struct{})
count := 0 // keeps track of retries
pErr := db.Txn(func(txn *client.Txn) *roachpb.Error {
if test.isolation == roachpb.SNAPSHOT {
if pErr := txn.SetIsolation(roachpb.SNAPSHOT); pErr != nil {
return pErr
}
}
txn.InternalSetPriority(txnPri)
count++
// Lay down the intent.
if pErr := txn.Put(key, "txn-value"); pErr != nil {
return pErr
}
// The wait group lets us pause txn until after the non-txn method has run once.
wg := sync.WaitGroup{}
// On the first true, send the non-txn put or get.
if count == 1 {
// We use a "notifying" sender here, which allows us to know exactly when the
// call has been processed; otherwise, we'd be dependent on timing.
sender.reset(&wg)
// We must try the non-txn put or get in a goroutine because
// it might have to retry and will only succeed immediately in
// the event we can push.
go func() {
var pErr *roachpb.Error
for i := 0; ; i++ {
if _, ok := test.args.(*roachpb.GetRequest); ok {
_, pErr = db.Get(key)
} else {
pErr = db.Put(key, "value")
}
if _, ok := pErr.GetDetail().(*roachpb.WriteIntentError); !ok {
break
}
}
close(doneCall)
if pErr != nil {
t.Fatalf("%d: expected success on non-txn call to %s; got %s", i, test.args.Method(), pErr)
}
}()
sender.wait()
}
return nil
})
if pErr != nil {
t.Fatalf("%d: expected success writing transactionally; got %s", i, pErr)
}
// Make sure non-txn put or get has finished.
<-doneCall
// Get the current value to verify whether the txn happened first.
gr, pErr := db.Get(key)
if pErr != nil {
t.Fatalf("%d: expected success getting %q: %s", i, key, pErr)
}
if _, isGet := test.args.(*roachpb.GetRequest); isGet || test.canPush {
if !bytes.Equal(gr.ValueBytes(), []byte("txn-value")) {
t.Errorf("%d: expected \"txn-value\"; got %q", i, gr.ValueBytes())
}
} else {
if !bytes.Equal(gr.ValueBytes(), []byte("value")) {
t.Errorf("%d: expected \"value\"; got %q", i, gr.ValueBytes())
}
}
if count != test.expAttempts {
t.Errorf("%d: expected %d attempt(s); got %d", i, test.expAttempts, count)
}
}
}
// Batch implements the roachpb.KVServer interface.
func (n *Node) Batch(ctx context.Context, args *roachpb.BatchRequest) (*roachpb.BatchResponse, error) {
// TODO(marc): this code is duplicated in kv/db.go, which should be fixed.
// Also, grpc's authentication model (which gives credential access in the
// request handler) doesn't really fit with the current design of the
// security package (which assumes that TLS state is only given at connection
// time) - that should be fixed.
if peer, ok := peer.FromContext(ctx); ok {
if tlsInfo, ok := peer.AuthInfo.(credentials.TLSInfo); ok {
certUser, err := security.GetCertificateUser(&tlsInfo.State)
if err != nil {
return nil, err
}
if certUser != security.NodeUser {
return nil, util.Errorf("user %s is not allowed", certUser)
}
}
}
var br *roachpb.BatchResponse
opName := "node " + strconv.Itoa(int(n.Descriptor.NodeID)) // could save allocs here
fail := func(err error) {
br = &roachpb.BatchResponse{}
br.Error = roachpb.NewError(err)
}
f := func() {
sp, err := tracing.JoinOrNew(n.ctx.Tracer, args.Trace, opName)
if err != nil {
fail(err)
return
}
// If this is a snowball span, it gets special treatment: It skips the
// regular tracing machinery, and we instead send the collected spans
// back with the response. This is more expensive, but then again,
// those are individual requests traced by users, so they can be.
if sp.BaggageItem(tracing.Snowball) != "" {
sp.LogEvent("delegating to snowball tracing")
sp.Finish()
if sp, err = tracing.JoinOrNewSnowball(opName, args.Trace, func(rawSpan basictracer.RawSpan) {
encSp, err := tracing.EncodeRawSpan(&rawSpan, nil)
if err != nil {
log.Warning(err)
}
br.CollectedSpans = append(br.CollectedSpans, encSp)
}); err != nil {
fail(err)
return
}
}
defer sp.Finish()
traceCtx := opentracing.ContextWithSpan(n.context(ctx), sp)
tStart := timeutil.Now()
var pErr *roachpb.Error
br, pErr = n.stores.Send(traceCtx, *args)
if pErr != nil {
br = &roachpb.BatchResponse{}
log.Trace(traceCtx, fmt.Sprintf("error: %T", pErr.GetDetail()))
}
if br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(n.stores, br))
}
n.metrics.callComplete(timeutil.Since(tStart), pErr)
br.Error = pErr
}
if !n.stopper.RunTask(f) {
return nil, util.Errorf("node %d stopped", n.Descriptor.NodeID)
}
return br, nil
}
// sendChunk is in charge of sending an "admissible" piece of batch, i.e. one
// which doesn't need to be subdivided further before going to a range (so no
// mixing of forward and reverse scans, etc). The parameters and return values
// correspond to client.Sender with the exception of the returned boolean,
// which is true when indicating that the caller should retry but needs to send
// EndTransaction in a separate request.
func (ds *DistSender) sendChunk(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error, bool) {
isReverse := ba.IsReverse()
ctx, cleanup := tracing.EnsureContext(ctx, ds.Tracer)
defer cleanup()
// The minimal key range encompassing all requests contained within.
// Local addressing has already been resolved.
// TODO(tschottdorf): consider rudimentary validation of the batch here
// (for example, non-range requests with EndKey, or empty key ranges).
rs, err := keys.Range(ba)
if err != nil {
return nil, roachpb.NewError(err), false
}
var br *roachpb.BatchResponse
// Send the request to one range per iteration.
for {
// Increase the sequence counter only once before sending RPCs to
// the ranges involved in this chunk of the batch (as opposed to for
// each RPC individually). On RPC errors, there's no guarantee that
// the request hasn't made its way to the target regardless of the
// error; we'd like the second execution to be caught by the sequence
// cache if that happens. There is a small chance that that we address
// a range twice in this chunk (stale/suboptimal descriptors due to
// splits/merges) which leads to a transaction retry.
// TODO(tschottdorf): it's possible that if we don't evict from the
// cache we could be in for a busy loop.
ba.SetNewRequest()
var curReply *roachpb.BatchResponse
var desc *roachpb.RangeDescriptor
var evictToken evictionToken
var needAnother bool
var pErr *roachpb.Error
var finished bool
for r := retry.Start(ds.rpcRetryOptions); r.Next(); {
// Get range descriptor (or, when spanning range, descriptors). Our
// error handling below may clear them on certain errors, so we
// refresh (likely from the cache) on every retry.
log.Trace(ctx, "meta descriptor lookup")
desc, needAnother, evictToken, pErr = ds.getDescriptors(rs, evictToken, isReverse)
// getDescriptors may fail retryably if the first range isn't
// available via Gossip.
if pErr != nil {
log.Trace(ctx, "range descriptor lookup failed: "+pErr.String())
if pErr.Retryable {
if log.V(1) {
log.Warning(pErr)
}
continue
}
break
} else {
log.Trace(ctx, "looked up range descriptor")
}
if needAnother && br == nil {
// TODO(tschottdorf): we should have a mechanism for discovering
// range merges (descriptor staleness will mostly go unnoticed),
// or we'll be turning single-range queries into multi-range
// queries for no good reason.
// If there's no transaction and op spans ranges, possibly
// re-run as part of a transaction for consistency. The
// case where we don't need to re-run is if the read
// consistency is not required.
if ba.Txn == nil && ba.IsPossibleTransaction() &&
ba.ReadConsistency != roachpb.INCONSISTENT {
return nil, roachpb.NewError(&roachpb.OpRequiresTxnError{}), false
}
// If the request is more than but ends with EndTransaction, we
// want the caller to come again with the EndTransaction in an
// extra call.
if l := len(ba.Requests) - 1; l > 0 && ba.Requests[l].GetInner().Method() == roachpb.EndTransaction {
return nil, roachpb.NewError(errors.New("cannot send 1PC txn to multiple ranges")), true /* shouldSplitET */
}
}
// It's possible that the returned descriptor misses parts of the
// keys it's supposed to scan after it's truncated to match the
// descriptor. Example revscan [a,g), first desc lookup for "g"
// returns descriptor [c,d) -> [d,g) is never scanned.
// We evict and retry in such a case.
includesFrontOfCurSpan := func(rd *roachpb.RangeDescriptor) bool {
if isReverse {
// This approach is needed because rs.EndKey is exclusive.
return desc.ContainsKeyRange(desc.StartKey, rs.EndKey)
}
return desc.ContainsKey(rs.Key)
}
if !includesFrontOfCurSpan(desc) {
if err := evictToken.Evict(); err != nil {
return nil, roachpb.NewError(err), false
}
// On addressing errors, don't backoff; retry immediately.
r.Reset()
continue
}
curReply, pErr = func() (*roachpb.BatchResponse, *roachpb.Error) {
// Truncate the request to our current key range.
intersected, iErr := rs.Intersect(desc)
if iErr != nil {
return nil, roachpb.NewError(iErr)
}
truncBA, numActive, trErr := truncate(ba, intersected)
if numActive == 0 && trErr == nil {
// This shouldn't happen in the wild, but some tests
// exercise it.
return nil, roachpb.NewErrorf("truncation resulted in empty batch on [%s,%s): %s",
rs.Key, rs.EndKey, ba)
}
if trErr != nil {
return nil, roachpb.NewError(trErr)
}
return ds.sendSingleRange(ctx, truncBA, desc)
}()
// If sending succeeded, break this loop.
if pErr == nil {
finished = true
break
}
if log.V(1) {
log.Warningf("failed to invoke %s: %s", ba, pErr)
}
log.Trace(ctx, fmt.Sprintf("reply error: %T", pErr.GetDetail()))
// Error handling below.
// If retryable, allow retry. For range not found or range
// key mismatch errors, we don't backoff on the retry,
// but reset the backoff loop so we can retry immediately.
switch tErr := pErr.GetDetail().(type) {
case *roachpb.SendError:
// For an RPC error to occur, we must've been unable to contact
// any replicas. In this case, likely all nodes are down (or
// not getting back to us within a reasonable amount of time).
// We may simply not be trying to talk to the up-to-date
// replicas, so clearing the descriptor here should be a good
// idea.
if err := evictToken.Evict(); err != nil {
return nil, roachpb.NewError(err), false
}
if tErr.CanRetry() {
continue
}
case *roachpb.RangeNotFoundError:
// Range descriptor might be out of date - evict it. This is
// likely the result of a rebalance.
if err := evictToken.Evict(); err != nil {
return nil, roachpb.NewError(err), false
}
// On addressing errors, don't backoff; retry immediately.
r.Reset()
if log.V(1) {
log.Warning(tErr)
}
continue
case *roachpb.RangeKeyMismatchError:
// Range descriptor might be out of date - evict it. This is
// likely the result of a range split. If we have new range
// descriptors, insert them instead as long as they are different
// from the last descriptor to avoid endless loops.
var replacements []roachpb.RangeDescriptor
different := func(rd *roachpb.RangeDescriptor) bool {
return !desc.RSpan().Equal(rd.RSpan())
}
if tErr.MismatchedRange != nil && different(tErr.MismatchedRange) {
replacements = append(replacements, *tErr.MismatchedRange)
}
if tErr.SuggestedRange != nil && different(tErr.SuggestedRange) {
if includesFrontOfCurSpan(tErr.SuggestedRange) {
replacements = append(replacements, *tErr.SuggestedRange)
}
}
// Same as Evict() if replacements is empty.
if err := evictToken.EvictAndReplace(replacements...); err != nil {
return nil, roachpb.NewError(err), false
}
// On addressing errors, don't backoff; retry immediately.
r.Reset()
if log.V(1) {
log.Warning(tErr)
}
continue
case *roachpb.NotLeaderError:
newLeader := tErr.Leader
if newLeader != nil {
// Verify that leader is a known replica according to the
// descriptor. If not, we've got a stale range descriptor;
// evict cache.
if i, _ := desc.FindReplica(newLeader.StoreID); i == -1 {
if log.V(1) {
log.Infof("error indicates unknown leader %s, expunging descriptor %s", newLeader, desc)
}
if err := evictToken.Evict(); err != nil {
return nil, roachpb.NewError(err), false
}
}
} else {
// If the new leader is unknown, we were talking to a
// replica that is partitioned away from the majority. Our
// range descriptor may be stale, so clear the cache.
//
// TODO(bdarnell): An unknown-leader error doesn't
// necessarily mean our descriptor is stale. Ideally we
// would treat these errors more like SendError: retry on
// another node (at a lower level), and then if it reaches
// this level then we know we've exhausted our options and
// must clear the cache.
if err := evictToken.Evict(); err != nil {
return nil, roachpb.NewError(err), false
}
newLeader = &roachpb.ReplicaDescriptor{}
}
// Next, cache the new leader.
ds.updateLeaderCache(roachpb.RangeID(desc.RangeID), *newLeader)
if log.V(1) {
log.Warning(tErr)
}
r.Reset()
continue
case retry.Retryable:
if tErr.CanRetry() {
if log.V(1) {
log.Warning(tErr)
}
continue
}
}
break
}
// Immediately return if querying a range failed non-retryably.
if pErr != nil {
return nil, pErr, false
} else if !finished {
select {
case <-ds.rpcRetryOptions.Closer:
return nil, roachpb.NewError(&roachpb.NodeUnavailableError{}), false
default:
log.Fatal("exited retry loop with nil error but finished=false")
}
}
ba.Txn.Update(curReply.Txn)
if br == nil {
// First response from a Range.
br = curReply
} else {
// This was the second or later call in a cross-Range request.
// Combine the new response with the existing one.
if err := br.Combine(curReply); err != nil {
return nil, roachpb.NewError(err), false
}
}
if ba.MaxScanResults > 0 {
// Count how many results we received.
var numResults int64
for _, resp := range curReply.Responses {
if cResp, ok := resp.GetInner().(roachpb.Countable); ok {
numResults += cResp.Count()
}
}
if numResults > ba.MaxScanResults {
panic(fmt.Sprintf("received %d results, limit was %d", numResults, ba.MaxScanResults))
}
ba.MaxScanResults -= numResults
if ba.MaxScanResults == 0 {
// We are done with this batch. Some requests might have NoopResponses; we must
// replace them with empty responses of the proper type.
for i, req := range ba.Requests {
if _, ok := br.Responses[i].GetInner().(*roachpb.NoopResponse); !ok {
continue
}
union := roachpb.ResponseUnion{}
var reply roachpb.Response
if _, ok := req.GetInner().(*roachpb.ScanRequest); ok {
reply = &roachpb.ScanResponse{}
} else {
_ = req.GetInner().(*roachpb.ReverseScanRequest)
reply = &roachpb.ReverseScanResponse{}
}
union.MustSetInner(reply)
br.Responses[i] = union
}
return br, nil, false
}
}
// If this request has a bound (such as MaxResults in
// ScanRequest) and we are going to query at least one more range,
// check whether enough rows have been retrieved.
// TODO(tschottdorf): need tests for executing a multi-range batch
// with various bounded requests which saturate at different times.
if needAnother {
// Start with the assumption that all requests are saturated.
// Below, we look at each and decide whether that's true.
// Everything that is indeed saturated is "masked out" from the
// batch request; only if that's all requests does needAnother
// remain false.
needAnother = false
if br == nil {
// Clone ba.Requests. This is because we're multi-range, and
// some requests may be bounded, which could lead to them being
// masked out once they're saturated. We don't want to risk
// removing requests that way in the "master copy" since that
// could lead to omitting requests in certain retry scenarios.
ba.Requests = append([]roachpb.RequestUnion(nil), ba.Requests...)
}
for i, union := range ba.Requests {
args := union.GetInner()
if _, ok := args.(*roachpb.NoopRequest); ok {
// NoopRequests are skipped.
continue
}
boundedArg, ok := args.(roachpb.Bounded)
if !ok {
// Non-bounded request. We will have to query all ranges.
needAnother = true
continue
}
prevBound := boundedArg.GetBound()
cReply, ok := curReply.Responses[i].GetInner().(roachpb.Countable)
if !ok || prevBound <= 0 {
// Request bounded, but without max results. Again, will
// need to query everything we can. The case in which the reply
// isn't countable occurs when the request wasn't active for
// that range (since it didn't apply to it), so the response
// is a NoopResponse.
needAnother = true
continue
}
nextBound := prevBound - cReply.Count()
if nextBound <= 0 {
// We've hit max results for this piece of the batch. Mask
// it out (we've copied the requests slice above, so this
// is kosher).
union := &ba.Requests[i] // avoid working on copy
union.MustSetInner(&noopRequest)
continue
}
// The request isn't saturated yet.
needAnother = true
boundedArg.SetBound(nextBound)
}
}
// If this was the last range accessed by this call, exit loop.
if !needAnother {
return br, nil, false
}
if isReverse {
// In next iteration, query previous range.
// We use the StartKey of the current descriptor as opposed to the
// EndKey of the previous one since that doesn't have bugs when
// stale descriptors come into play.
rs.EndKey, err = prev(ba, desc.StartKey)
} else {
// In next iteration, query next range.
// It's important that we use the EndKey of the current descriptor
// as opposed to the StartKey of the next one: if the former is stale,
// it's possible that the next range has since merged the subsequent
// one, and unless both descriptors are stale, the next descriptor's
// StartKey would move us to the beginning of the current range,
// resulting in a duplicate scan.
rs.Key, err = next(ba, desc.EndKey)
}
if err != nil {
return nil, roachpb.NewError(err), false
}
log.Trace(ctx, "querying next range")
}
}
// updateState updates the transaction state in both the success and
// error cases, applying those updates to the corresponding txnMeta
// object when adequate. It also updates certain errors with the
// updated transaction for use by client restarts.
func (tc *TxnCoordSender) updateState(ctx context.Context, ba roachpb.BatchRequest, br *roachpb.BatchResponse, pErr *roachpb.Error) *roachpb.Error {
trace := tracer.FromCtx(ctx)
newTxn := &roachpb.Transaction{}
newTxn.Update(ba.Txn)
// TODO(tamird): remove this clone. It's currently needed to avoid race conditions.
pErr = proto.Clone(pErr).(*roachpb.Error)
// TODO(bdarnell): We're writing to errors here (and where using ErrorWithIndex);
// since there's no concept of ownership copy-on-write is always preferable.
switch t := pErr.GetDetail().(type) {
case nil:
newTxn.Update(br.Txn)
// Move txn timestamp forward to response timestamp if applicable.
// TODO(tschottdorf): see (*Replica).executeBatch and comments within.
// Looks like this isn't necessary any more, nor did it prevent a bug
// referenced in a TODO there.
newTxn.Timestamp.Forward(br.Timestamp)
case *roachpb.TransactionStatusError:
// Likely already committed or more obscure errors such as epoch or
// timestamp regressions; consider txn dead.
defer tc.cleanupTxn(trace, *pErr.GetTxn())
case *roachpb.OpRequiresTxnError:
panic("OpRequiresTxnError must not happen at this level")
case *roachpb.ReadWithinUncertaintyIntervalError:
// Mark the host as certain. See the protobuf comment for
// Transaction.CertainNodes for details.
if t.NodeID == 0 {
panic("no replica set in header on uncertainty restart")
}
newTxn.Update(pErr.GetTxn())
newTxn.CertainNodes.Add(t.NodeID)
// If the reader encountered a newer write within the uncertainty
// interval, move the timestamp forward, just past that write or
// up to MaxTimestamp, whichever comes first.
candidateTS := newTxn.MaxTimestamp
candidateTS.Backward(t.ExistingTimestamp.Add(0, 1))
newTxn.Timestamp.Forward(candidateTS)
newTxn.Restart(ba.UserPriority, newTxn.Priority, newTxn.Timestamp)
pErr.SetTxn(newTxn)
case *roachpb.TransactionAbortedError:
trace.SetError()
newTxn.Update(pErr.GetTxn())
// Increase timestamp if applicable.
newTxn.Timestamp.Forward(pErr.GetTxn().Timestamp)
newTxn.Priority = pErr.GetTxn().Priority
pErr.SetTxn(newTxn)
// Clean up the freshly aborted transaction in defer(), avoiding a
// race with the state update below.
defer tc.cleanupTxn(trace, *pErr.GetTxn())
case *roachpb.TransactionPushError:
newTxn.Update(t.Txn)
// Increase timestamp if applicable, ensuring that we're
// just ahead of the pushee.
newTxn.Timestamp.Forward(t.PusheeTxn.Timestamp.Add(0, 1))
newTxn.Restart(ba.UserPriority, t.PusheeTxn.Priority-1, newTxn.Timestamp)
t.Txn = newTxn
pErr.SetTxn(newTxn)
case *roachpb.TransactionRetryError:
newTxn.Update(pErr.GetTxn())
newTxn.Restart(ba.UserPriority, pErr.GetTxn().Priority, newTxn.Timestamp)
pErr.SetTxn(newTxn)
default:
trace.SetError()
}
return func() *roachpb.Error {
if len(newTxn.ID) <= 0 {
return pErr
}
id := string(newTxn.ID)
tc.Lock()
defer tc.Unlock()
txnMeta := tc.txns[id]
// For successful transactional requests, keep the written intents and
// the updated transaction record to be sent along with the reply.
// The transaction metadata is created with the first writing operation.
// A tricky edge case is that of a transaction which "fails" on the
// first writing request, but actually manages to write some intents
// (for example, due to being multi-range). In this case, there will
// be an error, but the transaction will be marked as Writing and the
// coordinator must track the state, for the client's retry will be
// performed with a Writing transaction which the coordinator rejects
// unless it is tracking it (on top of it making sense to track it;
// after all, it **has** laid down intents and only the coordinator
// can augment a potential EndTransaction call). See #3303.
intents := ba.GetIntentSpans()
if len(intents) > 0 && (pErr == nil || newTxn.Writing) {
if txnMeta == nil {
if !newTxn.Writing {
panic("txn with intents marked as non-writing")
}
// If the transaction is already over, there's no point in
// launching a one-off coordinator which will shut down right
// away. If we ended up here with an error, we'll always start
// the coordinator - the transaction has laid down intents, so
// we expect it to be committed/aborted at some point in the
// future.
if _, isEnding := ba.GetArg(roachpb.EndTransaction); pErr != nil || !isEnding {
trace.Event("coordinator spawns")
txnMeta = &txnMetadata{
txn: *newTxn,
keys: cache.NewIntervalCache(cache.Config{Policy: cache.CacheNone}),
firstUpdateNanos: tc.clock.PhysicalNow(),
lastUpdateNanos: tc.clock.PhysicalNow(),
timeoutDuration: tc.clientTimeout,
txnEnd: make(chan struct{}),
}
tc.txns[id] = txnMeta
if !tc.stopper.RunAsyncTask(func() {
tc.heartbeatLoop(id)
}) {
// The system is already draining and we can't start the
// heartbeat. We refuse new transactions for now because
// they're likely not going to have all intents committed.
// In principle, we can relax this as needed though.
tc.unregisterTxnLocked(id)
return roachpb.NewError(&roachpb.NodeUnavailableError{})
}
}
}
}
// Update our record of this transaction, even on error.
if txnMeta != nil {
txnMeta.txn = *newTxn
if !txnMeta.txn.Writing {
panic("tracking a non-writing txn")
}
txnMeta.setLastUpdate(tc.clock.PhysicalNow())
// Adding the intents even on error reduces the likelyhood of dangling
// intents blocking concurrent writers for extended periods of time.
// See #3346.
for _, intent := range intents {
txnMeta.addKeyRange(intent.Key, intent.EndKey)
}
}
if pErr == nil {
// For successful transactional requests, always send the updated txn
// record back.
br.Txn = newTxn
}
return pErr
}()
}
// Send implements the batch.Sender interface. If the request is part of a
// transaction, the TxnCoordSender adds the transaction to a map of active
// transactions and begins heartbeating it. Every subsequent request for the
// same transaction updates the lastUpdate timestamp to prevent live
// transactions from being considered abandoned and garbage collected.
// Read/write mutating requests have their key or key range added to the
// transaction's interval tree of key ranges for eventual cleanup via resolved
// write intents; they're tagged to an outgoing EndTransaction request, with
// the receiving replica in charge of resolving them.
func (tc *TxnCoordSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if err := tc.maybeBeginTxn(&ba); err != nil {
return nil, roachpb.NewError(err)
}
var startNS int64
ba.SetNewRequest()
// This is the earliest point at which the request has an ID (if
// applicable). Begin a Trace which follows this request.
trace := tc.tracer.NewTrace(tracer.Coord, &ba)
defer trace.Finalize()
defer trace.Epoch("sending batch")()
ctx = tracer.ToCtx(ctx, trace)
var id string // optional transaction ID
if ba.Txn != nil {
// If this request is part of a transaction...
id = string(ba.Txn.ID)
// Verify that if this Transaction is not read-only, we have it on
// file. If not, refuse writes - the client must have issued a write on
// another coordinator previously.
if ba.Txn.Writing && ba.IsTransactionWrite() {
tc.Lock()
_, ok := tc.txns[id]
tc.Unlock()
if !ok {
return nil, roachpb.NewErrorf("transaction must not write on multiple coordinators")
}
}
// Set the timestamp to the original timestamp for read-only
// commands and to the transaction timestamp for read/write
// commands.
if ba.IsReadOnly() {
ba.Timestamp = ba.Txn.OrigTimestamp
} else {
ba.Timestamp = ba.Txn.Timestamp
}
if rArgs, ok := ba.GetArg(roachpb.EndTransaction); ok {
et := rArgs.(*roachpb.EndTransactionRequest)
if len(et.Key) != 0 {
return nil, roachpb.NewErrorf("EndTransaction must not have a Key set")
}
et.Key = ba.Txn.Key
// Remember when EndTransaction started in case we want to
// be linearizable.
startNS = tc.clock.PhysicalNow()
if len(et.IntentSpans) > 0 {
// TODO(tschottdorf): it may be useful to allow this later.
// That would be part of a possible plan to allow txns which
// write on multiple coordinators.
return nil, roachpb.NewErrorf("client must not pass intents to EndTransaction")
}
tc.Lock()
txnMeta, metaOK := tc.txns[id]
if id != "" && metaOK {
et.IntentSpans = txnMeta.intentSpans()
}
tc.Unlock()
if intentSpans := ba.GetIntentSpans(); len(intentSpans) > 0 {
// Writes in Batch, so EndTransaction is fine. Should add
// outstanding intents to EndTransaction, though.
// TODO(tschottdorf): possible issues when the batch fails,
// but the intents have been added anyways.
// TODO(tschottdorf): some of these intents may be covered
// by others, for example {[a,b), a}). This can lead to
// some extra requests when those are non-local to the txn
// record. But it doesn't seem worth optimizing now.
et.IntentSpans = append(et.IntentSpans, intentSpans...)
} else if !metaOK {
// If we don't have the transaction, then this must be a retry
// by the client. We can no longer reconstruct a correct
// request so we must fail.
//
// TODO(bdarnell): if we had a GetTransactionStatus API then
// we could lookup the transaction and return either nil or
// TransactionAbortedError instead of this ambivalent error.
return nil, roachpb.NewErrorf("transaction is already committed or aborted")
}
if len(et.IntentSpans) == 0 {
// If there aren't any intents, then there's factually no
// transaction to end. Read-only txns have all of their state in
// the client.
return nil, roachpb.NewErrorf("cannot commit a read-only transaction")
}
if log.V(1) {
for _, intent := range et.IntentSpans {
trace.Event(fmt.Sprintf("intent: [%s,%s)", intent.Key, intent.EndKey))
}
}
}
}
// Send the command through wrapped sender, taking appropriate measures
// on error.
var br *roachpb.BatchResponse
{
var pErr *roachpb.Error
br, pErr = tc.wrapped.Send(ctx, ba)
if _, ok := pErr.GetDetail().(*roachpb.OpRequiresTxnError); ok {
br, pErr = tc.resendWithTxn(ba)
}
if pErr = tc.updateState(ctx, ba, br, pErr); pErr != nil {
trace.Event(fmt.Sprintf("error: %s", pErr))
return nil, pErr
}
}
if br.Txn == nil {
return br, nil
}
if _, ok := ba.GetArg(roachpb.EndTransaction); !ok {
return br, nil
}
// If the --linearizable flag is set, we want to make sure that
// all the clocks in the system are past the commit timestamp
// of the transaction. This is guaranteed if either
// - the commit timestamp is MaxOffset behind startNS
// - MaxOffset ns were spent in this function
// when returning to the client. Below we choose the option
// that involves less waiting, which is likely the first one
// unless a transaction commits with an odd timestamp.
if tsNS := br.Txn.Timestamp.WallTime; startNS > tsNS {
startNS = tsNS
}
sleepNS := tc.clock.MaxOffset() -
time.Duration(tc.clock.PhysicalNow()-startNS)
if tc.linearizable && sleepNS > 0 {
defer func() {
if log.V(1) {
log.Infof("%v: waiting %s on EndTransaction for linearizability", br.Txn.Short(), util.TruncateDuration(sleepNS, time.Millisecond))
}
time.Sleep(sleepNS)
}()
}
if br.Txn.Status != roachpb.PENDING {
tc.cleanupTxn(trace, *br.Txn)
}
return br, nil
}
// sendChunk is in charge of sending an "admissible" piece of batch, i.e. one
// which doesn't need to be subdivided further before going to a range (so no
// mixing of forward and reverse scans, etc). The parameters and return values
// correspond to client.Sender with the exception of the returned boolean,
// which is true when indicating that the caller should retry but needs to send
// EndTransaction in a separate request.
func (ds *DistSender) sendChunk(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error, bool) {
isReverse := ba.IsReverse()
// TODO(radu): when contexts are properly plumbed, we should be able to get
// the tracer from ctx, not from the DistSender.
ctx, cleanup := tracing.EnsureContext(ctx, tracing.TracerFromCtx(ds.Ctx))
defer cleanup()
// The minimal key range encompassing all requests contained within.
// Local addressing has already been resolved.
// TODO(tschottdorf): consider rudimentary validation of the batch here
// (for example, non-range requests with EndKey, or empty key ranges).
rs, err := keys.Range(ba)
if err != nil {
return nil, roachpb.NewError(err), false
}
var br *roachpb.BatchResponse
// Send the request to one range per iteration.
for {
// Increase the sequence counter only once before sending RPCs to
// the ranges involved in this chunk of the batch (as opposed to for
// each RPC individually). On RPC errors, there's no guarantee that
// the request hasn't made its way to the target regardless of the
// error; we'd like the second execution to be caught by the sequence
// cache if that happens. There is a small chance that that we address
// a range twice in this chunk (stale/suboptimal descriptors due to
// splits/merges) which leads to a transaction retry.
// TODO(tschottdorf): it's possible that if we don't evict from the
// cache we could be in for a busy loop.
ba.SetNewRequest()
var curReply *roachpb.BatchResponse
var desc *roachpb.RangeDescriptor
var evictToken *evictionToken
var needAnother bool
var pErr *roachpb.Error
var finished bool
var numAttempts int
for r := retry.StartWithCtx(ctx, ds.rpcRetryOptions); r.Next(); {
numAttempts++
{
const magicLogCurAttempt = 20
var seq int32
if ba.Txn != nil {
seq = ba.Txn.Sequence
}
if numAttempts%magicLogCurAttempt == 0 || seq%magicLogCurAttempt == 0 {
// Log a message if a request appears to get stuck for a long
// time or, potentially, forever. See #8975.
// The local counter captures this loop here; the Sequence number
// should capture anything higher up (as it needs to be
// incremented every time this method is called).
log.Warningf(
ctx,
"%d retries for an RPC at sequence %d, last error was: %s, remaining key ranges %s: %s",
numAttempts, seq, pErr, rs, ba,
)
}
}
// Get range descriptor (or, when spanning range, descriptors). Our
// error handling below may clear them on certain errors, so we
// refresh (likely from the cache) on every retry.
log.Trace(ctx, "meta descriptor lookup")
var err error
desc, needAnother, evictToken, err = ds.getDescriptors(ctx, rs, evictToken, isReverse)
// getDescriptors may fail retryably if, for example, the first
// range isn't available via Gossip. Assume that all errors at
// this level are retryable. Non-retryable errors would be for
// things like malformed requests which we should have checked
// for before reaching this point.
if err != nil {
log.Trace(ctx, "range descriptor lookup failed: "+err.Error())
if log.V(1) {
log.Warning(ctx, err)
}
pErr = roachpb.NewError(err)
continue
}
if needAnother && br == nil {
// TODO(tschottdorf): we should have a mechanism for discovering
// range merges (descriptor staleness will mostly go unnoticed),
// or we'll be turning single-range queries into multi-range
// queries for no good reason.
// If there's no transaction and op spans ranges, possibly
// re-run as part of a transaction for consistency. The
// case where we don't need to re-run is if the read
// consistency is not required.
if ba.Txn == nil && ba.IsPossibleTransaction() &&
ba.ReadConsistency != roachpb.INCONSISTENT {
return nil, roachpb.NewError(&roachpb.OpRequiresTxnError{}), false
}
// If the request is more than but ends with EndTransaction, we
// want the caller to come again with the EndTransaction in an
// extra call.
if l := len(ba.Requests) - 1; l > 0 && ba.Requests[l].GetInner().Method() == roachpb.EndTransaction {
return nil, roachpb.NewError(errors.New("cannot send 1PC txn to multiple ranges")), true /* shouldSplitET */
}
}
// It's possible that the returned descriptor misses parts of the
// keys it's supposed to scan after it's truncated to match the
// descriptor. Example revscan [a,g), first desc lookup for "g"
// returns descriptor [c,d) -> [d,g) is never scanned.
// We evict and retry in such a case.
includesFrontOfCurSpan := func(rd *roachpb.RangeDescriptor) bool {
if isReverse {
return desc.ContainsExclusiveEndKey(rs.EndKey)
}
return desc.ContainsKey(rs.Key)
}
if !includesFrontOfCurSpan(desc) {
if err := evictToken.Evict(ctx); err != nil {
return nil, roachpb.NewError(err), false
}
// On addressing errors, don't backoff; retry immediately.
r.Reset()
continue
}
curReply, pErr = func() (*roachpb.BatchResponse, *roachpb.Error) {
// Truncate the request to our current key range.
intersected, iErr := rs.Intersect(desc)
if iErr != nil {
return nil, roachpb.NewError(iErr)
}
truncBA, numActive, trErr := truncate(ba, intersected)
if numActive == 0 && trErr == nil {
// This shouldn't happen in the wild, but some tests
// exercise it.
return nil, roachpb.NewErrorf("truncation resulted in empty batch on [%s,%s): %s",
rs.Key, rs.EndKey, ba)
}
if trErr != nil {
return nil, roachpb.NewError(trErr)
}
return ds.sendSingleRange(ctx, truncBA, desc)
}()
// If sending succeeded, break this loop.
if pErr == nil {
finished = true
break
}
log.VTracef(1, ctx, "reply error %s: %s", ba, pErr)
// Error handling: If the error indicates that our range
// descriptor is out of date, evict it from the cache and try
// again. Errors that apply only to a single replica were
// handled in send().
//
// TODO(bdarnell): Don't retry endlessly. If we fail twice in a
// row and the range descriptor hasn't changed, return the error
// to our caller.
switch tErr := pErr.GetDetail().(type) {
case *roachpb.SendError:
// We've tried all the replicas without success. Either
// they're all down, or we're using an out-of-date range
// descriptor. Invalidate the cache and try again with the new
// metadata.
if err := evictToken.Evict(ctx); err != nil {
return nil, roachpb.NewError(err), false
}
continue
case *roachpb.RangeKeyMismatchError:
// Range descriptor might be out of date - evict it. This is
// likely the result of a range split. If we have new range
// descriptors, insert them instead as long as they are different
// from the last descriptor to avoid endless loops.
var replacements []roachpb.RangeDescriptor
different := func(rd *roachpb.RangeDescriptor) bool {
return !desc.RSpan().Equal(rd.RSpan())
}
if tErr.MismatchedRange != nil && different(tErr.MismatchedRange) {
replacements = append(replacements, *tErr.MismatchedRange)
}
if tErr.SuggestedRange != nil && different(tErr.SuggestedRange) {
if includesFrontOfCurSpan(tErr.SuggestedRange) {
replacements = append(replacements, *tErr.SuggestedRange)
}
}
// Same as Evict() if replacements is empty.
if err := evictToken.EvictAndReplace(ctx, replacements...); err != nil {
return nil, roachpb.NewError(err), false
}
// On addressing errors, don't backoff; retry immediately.
r.Reset()
if log.V(1) {
log.Warning(ctx, tErr)
}
continue
}
break
}
// Immediately return if querying a range failed non-retryably.
if pErr != nil {
return nil, pErr, false
} else if !finished {
select {
case <-ds.rpcRetryOptions.Closer:
return nil, roachpb.NewError(&roachpb.NodeUnavailableError{}), false
case <-ctx.Done():
return nil, roachpb.NewError(ctx.Err()), false
default:
log.Fatal(ctx, "exited retry loop with nil error but finished=false")
}
}
ba.UpdateTxn(curReply.Txn)
if br == nil {
// First response from a Range.
br = curReply
} else {
// This was the second or later call in a cross-Range request.
// Combine the new response with the existing one.
if err := br.Combine(curReply); err != nil {
return nil, roachpb.NewError(err), false
}
}
if isReverse {
// In next iteration, query previous range.
// We use the StartKey of the current descriptor as opposed to the
// EndKey of the previous one since that doesn't have bugs when
// stale descriptors come into play.
rs.EndKey, err = prev(ba, desc.StartKey)
} else {
// In next iteration, query next range.
// It's important that we use the EndKey of the current descriptor
// as opposed to the StartKey of the next one: if the former is stale,
// it's possible that the next range has since merged the subsequent
// one, and unless both descriptors are stale, the next descriptor's
// StartKey would move us to the beginning of the current range,
// resulting in a duplicate scan.
rs.Key, err = next(ba, desc.EndKey)
}
if err != nil {
return nil, roachpb.NewError(err), false
}
if ba.MaxSpanRequestKeys > 0 {
// Count how many results we received.
var numResults int64
for _, resp := range curReply.Responses {
numResults += resp.GetInner().Header().NumKeys
}
if numResults > ba.MaxSpanRequestKeys {
panic(fmt.Sprintf("received %d results, limit was %d", numResults, ba.MaxSpanRequestKeys))
}
ba.MaxSpanRequestKeys -= numResults
if ba.MaxSpanRequestKeys == 0 {
// prepare the batch response after meeting the max key limit.
fillSkippedResponses(ba, br, rs)
// done, exit loop.
return br, nil, false
}
}
// If this was the last range accessed by this call, exit loop.
if !needAnother {
return br, nil, false
}
// key cannot be less that the end key.
if !rs.Key.Less(rs.EndKey) {
panic(fmt.Sprintf("start key %s is less than %s", rs.Key, rs.EndKey))
}
log.Trace(ctx, "querying next range")
}
}
// updateState updates the transaction state in both the success and
// error cases, applying those updates to the corresponding txnMeta
// object when adequate. It also updates certain errors with the
// updated transaction for use by client restarts.
func (tc *TxnCoordSender) updateState(ctx context.Context, ba roachpb.BatchRequest, br *roachpb.BatchResponse, pErr *roachpb.Error) *roachpb.Error {
sp, cleanupSp := tracing.SpanFromContext(opTxnCoordSender, tc.tracer, ctx)
defer cleanupSp()
newTxn := &roachpb.Transaction{}
newTxn.Update(ba.Txn)
if pErr == nil {
newTxn.Update(br.Txn)
} else {
newTxn.Update(pErr.GetTxn())
}
// If the request was successful but we're in a transaction which needs to
// restart but doesn't know it yet, let it restart now (as opposed to
// waiting until EndTransaction).
if pErr == nil && newTxn.Isolation == roachpb.SERIALIZABLE &&
!newTxn.OrigTimestamp.Equal(newTxn.Timestamp) {
pErr = roachpb.NewErrorWithTxn(roachpb.NewTransactionRetryError(), br.Txn)
br = nil
}
switch t := pErr.GetDetail().(type) {
case nil:
// Move txn timestamp forward to response timestamp if applicable.
// TODO(tschottdorf): see (*Replica).executeBatch and comments within.
// Looks like this isn't necessary any more, nor did it prevent a bug
// referenced in a TODO there.
newTxn.Timestamp.Forward(br.Timestamp)
case *roachpb.TransactionStatusError:
// Likely already committed or more obscure errors such as epoch or
// timestamp regressions; consider txn dead.
defer tc.cleanupTxn(sp, *pErr.GetTxn())
case *roachpb.OpRequiresTxnError:
panic("OpRequiresTxnError must not happen at this level")
case *roachpb.ReadWithinUncertaintyIntervalError:
// If the reader encountered a newer write within the uncertainty
// interval, we advance the txn's timestamp just past the last observed
// timestamp from the node.
restartTS, ok := newTxn.GetObservedTimestamp(pErr.OriginNode)
if !ok {
pErr = roachpb.NewError(util.Errorf("no observed timestamp for node %d found on uncertainty restart", pErr.OriginNode))
} else {
newTxn.Timestamp.Forward(restartTS)
newTxn.Restart(ba.UserPriority, newTxn.Priority, newTxn.Timestamp)
}
case *roachpb.TransactionAbortedError:
// Increase timestamp if applicable.
newTxn.Timestamp.Forward(pErr.GetTxn().Timestamp)
newTxn.Priority = pErr.GetTxn().Priority
// Clean up the freshly aborted transaction in defer(), avoiding a
// race with the state update below.
defer tc.cleanupTxn(sp, *newTxn)
case *roachpb.TransactionPushError:
// Increase timestamp if applicable, ensuring that we're
// just ahead of the pushee.
newTxn.Timestamp.Forward(t.PusheeTxn.Timestamp.Add(0, 1))
newTxn.Restart(ba.UserPriority, t.PusheeTxn.Priority-1, newTxn.Timestamp)
case *roachpb.TransactionRetryError:
newTxn.Restart(ba.UserPriority, pErr.GetTxn().Priority, newTxn.Timestamp)
default:
if pErr.GetTxn() != nil {
if pErr.CanRetry() {
panic("Retryable internal error must not happen at this level")
} else {
// Do not clean up the transaction here since the client might still
// want to continue the transaction. For example, a client might
// continue its transaction after receiving ConditionFailedError, which
// can come from a unique index violation.
}
}
}
if pErr != nil && pErr.GetTxn() != nil {
// Avoid changing existing errors because sometimes they escape into
// goroutines and then there are races. Fairly sure there isn't one
// here, but better safe than sorry.
pErrShallow := *pErr
pErrShallow.SetTxn(newTxn)
pErr = &pErrShallow
}
if newTxn.ID == nil {
return pErr
}
txnID := *newTxn.ID
tc.Lock()
defer tc.Unlock()
txnMeta := tc.txns[txnID]
// For successful transactional requests, keep the written intents and
// the updated transaction record to be sent along with the reply.
// The transaction metadata is created with the first writing operation.
// A tricky edge case is that of a transaction which "fails" on the
// first writing request, but actually manages to write some intents
// (for example, due to being multi-range). In this case, there will
// be an error, but the transaction will be marked as Writing and the
// coordinator must track the state, for the client's retry will be
// performed with a Writing transaction which the coordinator rejects
// unless it is tracking it (on top of it making sense to track it;
// after all, it **has** laid down intents and only the coordinator
// can augment a potential EndTransaction call). See #3303.
intents := ba.GetIntentSpans()
if len(intents) > 0 && (pErr == nil || newTxn.Writing) {
if txnMeta == nil {
if !newTxn.Writing {
panic("txn with intents marked as non-writing")
}
// If the transaction is already over, there's no point in
// launching a one-off coordinator which will shut down right
// away. If we ended up here with an error, we'll always start
// the coordinator - the transaction has laid down intents, so
// we expect it to be committed/aborted at some point in the
// future.
if _, isEnding := ba.GetArg(roachpb.EndTransaction); pErr != nil || !isEnding {
sp.LogEvent("coordinator spawns")
txnMeta = &txnMetadata{
txn: *newTxn,
keys: cache.NewIntervalCache(cache.Config{Policy: cache.CacheNone}),
firstUpdateNanos: tc.clock.PhysicalNow(),
lastUpdateNanos: tc.clock.PhysicalNow(),
timeoutDuration: tc.clientTimeout,
txnEnd: make(chan struct{}),
}
tc.txns[txnID] = txnMeta
if !tc.stopper.RunAsyncTask(func() {
tc.heartbeatLoop(txnID)
}) {
// The system is already draining and we can't start the
// heartbeat. We refuse new transactions for now because
// they're likely not going to have all intents committed.
// In principle, we can relax this as needed though.
tc.unregisterTxnLocked(txnID)
return roachpb.NewError(&roachpb.NodeUnavailableError{})
}
}
}
}
// Update our record of this transaction, even on error.
if txnMeta != nil {
txnMeta.txn = *newTxn
if !txnMeta.txn.Writing {
panic("tracking a non-writing txn")
}
txnMeta.setLastUpdate(tc.clock.PhysicalNow())
// Adding the intents even on error reduces the likelihood of dangling
// intents blocking concurrent writers for extended periods of time.
// See #3346.
for _, intent := range intents {
txnMeta.addKeyRange(intent.Key, intent.EndKey)
}
}
if pErr == nil {
// For successful transactional requests, always send the updated txn
// record back.
br.Txn = newTxn
}
return pErr
}