// TestTxnCoordSenderSingleRoundtripTxn checks that a batch which completely
// holds the writing portion of a Txn (including EndTransaction) does not
// launch a heartbeat goroutine at all.
func TestTxnCoordSenderSingleRoundtripTxn(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
manual := hlc.NewManualClock(123)
clock := hlc.NewClock(manual.UnixNano, 20*time.Nanosecond)
senderFunc := func(_ context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
br := ba.CreateReply()
txnClone := ba.Txn.Clone()
br.Txn = &txnClone
br.Txn.Writing = true
return br, nil
}
ambient := log.AmbientContext{Tracer: tracing.NewTracer()}
ts := NewTxnCoordSender(
ambient, senderFn(senderFunc), clock, false, stopper, MakeTxnMetrics(metric.TestSampleInterval),
)
// Stop the stopper manually, prior to trying the transaction. This has the
// effect of returning a NodeUnavailableError for any attempts at launching
// a heartbeat goroutine.
stopper.Stop()
var ba roachpb.BatchRequest
key := roachpb.Key("test")
ba.Add(&roachpb.BeginTransactionRequest{Span: roachpb.Span{Key: key}})
ba.Add(&roachpb.PutRequest{Span: roachpb.Span{Key: key}})
ba.Add(&roachpb.EndTransactionRequest{})
ba.Txn = &roachpb.Transaction{Name: "test"}
_, pErr := ts.Send(context.Background(), ba)
if pErr != nil {
t.Fatal(pErr)
}
}
// TestTxnCoordSenderErrorWithIntent validates that if a transactional request
// returns an error but also indicates a Writing transaction, the coordinator
// tracks it just like a successful request.
func TestTxnCoordSenderErrorWithIntent(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
defer stopper.Stop()
manual := hlc.NewManualClock(0)
clock := hlc.NewClock(manual.UnixNano)
clock.SetMaxOffset(20)
testCases := []struct {
roachpb.Error
errMsg string
}{
{*roachpb.NewError(roachpb.NewTransactionRetryError()), "retry txn"},
{*roachpb.NewError(roachpb.NewTransactionPushError(roachpb.Transaction{
TxnMeta: enginepb.TxnMeta{
ID: uuid.NewV4(),
}})), "failed to push"},
{*roachpb.NewErrorf("testError"), "testError"},
}
for i, test := range testCases {
func() {
senderFunc := func(_ context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
txn := ba.Txn.Clone()
txn.Writing = true
pErr := &roachpb.Error{}
*pErr = test.Error
pErr.SetTxn(&txn)
return nil, pErr
}
ambient := log.AmbientContext{Tracer: tracing.NewTracer()}
ts := NewTxnCoordSender(
ambient,
senderFn(senderFunc),
clock,
false,
stopper,
MakeTxnMetrics(metric.TestSampleInterval),
)
var ba roachpb.BatchRequest
key := roachpb.Key("test")
ba.Add(&roachpb.BeginTransactionRequest{Span: roachpb.Span{Key: key}})
ba.Add(&roachpb.PutRequest{Span: roachpb.Span{Key: key}})
ba.Add(&roachpb.EndTransactionRequest{})
ba.Txn = &roachpb.Transaction{Name: "test"}
_, pErr := ts.Send(context.Background(), ba)
if !testutils.IsPError(pErr, test.errMsg) {
t.Errorf("%d: error did not match %s: %v", i, test.errMsg, pErr)
}
defer teardownHeartbeats(ts)
ts.Lock()
defer ts.Unlock()
if len(ts.txns) != 1 {
t.Errorf("%d: expected transaction to be tracked", i)
}
}()
}
}
func TestBatchRequestString(t *testing.T) {
br := roachpb.BatchRequest{}
br.Txn = new(roachpb.Transaction)
for i := 0; i < 100; i++ {
br.Requests = append(br.Requests, roachpb.RequestUnion{Get: &roachpb.GetRequest{}})
}
br.Requests = append(br.Requests, roachpb.RequestUnion{EndTransaction: &roachpb.EndTransactionRequest{}})
e := `[txn: <nil>], Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), ... 76 skipped ..., Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), Get [/Min,/Min), EndTransaction [/Min,/Min)`
if e != br.String() {
t.Fatalf("e = %s, v = %s", e, br.String())
}
}
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(
ctx context.Context, ba roachpb.BatchRequest,
) (*roachpb.BatchResponse, *roachpb.Error) {
// If we aren't given a Replica, then a little bending over
// backwards here. This case applies exclusively to unittests.
if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
rs, err := keys.Range(ba)
if err != nil {
return nil, roachpb.NewError(err)
}
rangeID, repDesc, err := ls.LookupReplica(rs.Key, rs.EndKey)
if err != nil {
return nil, roachpb.NewError(err)
}
ba.RangeID = rangeID
ba.Replica = repDesc
}
store, err := ls.GetStore(ba.Replica.StoreID)
if err != nil {
return nil, roachpb.NewError(err)
}
if ba.Txn != nil {
// For calls that read data within a txn, we keep track of timestamps
// observed from the various participating nodes' HLC clocks. If we have
// a timestamp on file for this Node which is smaller than MaxTimestamp,
// we can lower MaxTimestamp accordingly. If MaxTimestamp drops below
// OrigTimestamp, we effectively can't see uncertainty restarts any
// more.
// Note that it's not an issue if MaxTimestamp propagates back out to
// the client via a returned Transaction update - when updating a Txn
// from another, the larger MaxTimestamp wins.
if maxTS, ok := ba.Txn.GetObservedTimestamp(ba.Replica.NodeID); ok && maxTS.Less(ba.Txn.MaxTimestamp) {
// Copy-on-write to protect others we might be sharing the Txn with.
shallowTxn := *ba.Txn
// The uncertainty window is [OrigTimestamp, maxTS), so if that window
// is empty, there won't be any uncertainty restarts.
if !ba.Txn.OrigTimestamp.Less(maxTS) {
log.Event(ctx, "read has no clock uncertainty")
}
shallowTxn.MaxTimestamp.Backward(maxTS)
ba.Txn = &shallowTxn
}
}
br, pErr := store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(store, br))
}
return br, pErr
}
// maybeBeginTxn begins a new transaction if a txn has been specified
// in the request but has a nil ID. The new transaction is initialized
// using the name and isolation in the otherwise uninitialized txn.
// The Priority, if non-zero is used as a minimum.
//
// No transactional writes are allowed unless preceded by a begin
// transaction request within the same batch. The exception is if the
// transaction is already in state txn.Writing=true.
func (tc *TxnCoordSender) maybeBeginTxn(ba *roachpb.BatchRequest) error {
if len(ba.Requests) == 0 {
return errors.Errorf("empty batch with txn")
}
if ba.Txn.ID == nil {
// Create transaction without a key. The key is set when a begin
// transaction request is received.
// The initial timestamp may be communicated by a higher layer.
// If so, use that. Otherwise make up a new one.
timestamp := ba.Txn.OrigTimestamp
if timestamp == hlc.ZeroTimestamp {
timestamp = tc.clock.Now()
}
newTxn := roachpb.NewTransaction(ba.Txn.Name, nil, ba.UserPriority,
ba.Txn.Isolation, timestamp, tc.clock.MaxOffset().Nanoseconds())
// Use existing priority as a minimum. This is used on transaction
// aborts to ratchet priority when creating successor transaction.
if newTxn.Priority < ba.Txn.Priority {
newTxn.Priority = ba.Txn.Priority
}
ba.Txn = newTxn
}
// Check for a begin transaction to set txn key based on the key of
// the first transactional write. Also enforce that no transactional
// writes occur before a begin transaction.
var haveBeginTxn bool
for _, req := range ba.Requests {
args := req.GetInner()
if bt, ok := args.(*roachpb.BeginTransactionRequest); ok {
if haveBeginTxn || ba.Txn.Writing {
return errors.Errorf("begin transaction requested twice in the same transaction: %s", ba.Txn)
}
haveBeginTxn = true
if ba.Txn.Key == nil {
ba.Txn.Key = bt.Key
}
}
if roachpb.IsTransactionWrite(args) && !haveBeginTxn && !ba.Txn.Writing {
return errors.Errorf("transactional write before begin transaction")
}
}
return nil
}
// tryAsyncAbort (synchronously) grabs a copy of the txn proto and the intents
// (which it then clears from txnMeta), and asynchronously tries to abort the
// transaction.
func (tc *TxnCoordSender) tryAsyncAbort(txnID uuid.UUID) {
tc.Lock()
txnMeta := tc.txns[txnID]
// Clone the intents and the txn to avoid data races.
intentSpans, _ := roachpb.MergeSpans(append([]roachpb.Span(nil), txnMeta.keys...))
txnMeta.keys = nil
txn := txnMeta.txn.Clone()
tc.Unlock()
// Since we don't hold the lock continuously, it's possible that two aborts
// raced here. That's fine (and probably better than the alternative, which
// is missing new intents sometimes).
if txn.Status != roachpb.PENDING {
return
}
ba := roachpb.BatchRequest{}
ba.Txn = &txn
et := &roachpb.EndTransactionRequest{
Span: roachpb.Span{
Key: txn.Key,
},
Commit: false,
IntentSpans: intentSpans,
}
ba.Add(et)
ctx := tc.AnnotateCtx(context.TODO())
if err := tc.stopper.RunAsyncTask(ctx, func(ctx context.Context) {
// Use the wrapped sender since the normal Sender does not allow
// clients to specify intents.
if _, pErr := tc.wrapped.Send(ctx, ba); pErr != nil {
if log.V(1) {
log.Warningf(ctx, "abort due to inactivity failed for %s: %s ", txn, pErr)
}
}
}); err != nil {
log.Warning(ctx, err)
}
}
// 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)
}
// divideAndSendBatchToRanges sends the supplied batch to all of the
// ranges which comprise the span specified by rs. The batch request
// is trimmed against each range which is part of the span and sent
// either serially or in parallel, if possible. isFirst indicates
// whether this is the first time this method has been called on the
// batch. It's specified false where this method is invoked recursively.
func (ds *DistSender) divideAndSendBatchToRanges(
ctx context.Context, ba roachpb.BatchRequest, rs roachpb.RSpan, isFirst bool,
) (br *roachpb.BatchResponse, pErr *roachpb.Error) {
// This function builds a channel of responses for each range
// implicated in the span (rs) and combines them into a single
// BatchResponse when finished.
var responseChs []chan response
defer func() {
for _, responseCh := range responseChs {
resp := <-responseCh
if resp.pErr != nil {
if pErr == nil {
pErr = resp.pErr
}
continue
}
if br == nil {
// First response from a Range.
br = resp.reply
} 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(resp.reply); err != nil {
pErr = roachpb.NewError(err)
return
}
br.Txn.Update(resp.reply.Txn)
}
}
// If we experienced an error, don't neglect to update the error's
// attached transaction with any responses which were received.
if pErr != nil {
if br != nil {
pErr.UpdateTxn(br.Txn)
}
}
}()
// Get initial seek key depending on direction of iteration.
var seekKey roachpb.RKey
isReverse := ba.IsReverse()
if isReverse {
seekKey = rs.EndKey
} else {
seekKey = rs.Key
}
// Send the request to one range per iteration.
ri := NewRangeIterator(ds, isReverse)
for ri.Seek(ctx, seekKey); ri.Valid(); ri.Seek(ctx, seekKey) {
// 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 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()
responseCh := make(chan response, 1)
responseChs = append(responseChs, responseCh)
if isFirst && ri.NeedAnother(rs) {
// 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 {
responseCh <- response{pErr: roachpb.NewError(&roachpb.OpRequiresTxnError{})}
return
}
// 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 {
responseCh <- response{pErr: errNo1PCTxn}
return
}
}
// Determine next seek key, taking a potentially sparse batch into
// consideration.
var err error
nextRS := rs
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.
seekKey, err = prev(ba, ri.Desc().StartKey)
nextRS.EndKey = seekKey
} 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.
seekKey, err = next(ba, ri.Desc().EndKey)
nextRS.Key = seekKey
}
if err != nil {
responseCh <- response{pErr: roachpb.NewError(err)}
return
}
// Send the next partial batch to the first range in the "rs" span.
// If we're not handling a request which limits responses and we
// can reserve one of the limited goroutines available for parallel
// batch RPCs, send asynchronously.
if ba.MaxSpanRequestKeys == 0 && ri.NeedAnother(rs) && ds.rpcContext != nil &&
ds.sendPartialBatchAsync(ctx, ba, rs, ri.Desc(), ri.Token(), isFirst, responseCh) {
// Note that we pass the batch request by value to the parallel
// goroutine to avoid using the cloned txn.
// Clone the txn to preserve the current txn sequence for the async call.
if ba.Txn != nil {
txnClone := ba.Txn.Clone()
ba.Txn = &txnClone
}
} else {
// Send synchronously if there is no parallel capacity left, there's a
// max results limit, or this is the final request in the span.
resp := ds.sendPartialBatch(ctx, ba, rs, ri.Desc(), ri.Token(), isFirst)
responseCh <- resp
if resp.pErr != nil {
return
}
ba.UpdateTxn(resp.reply.Txn)
// Check whether we've received enough responses to exit query loop.
if ba.MaxSpanRequestKeys > 0 {
var numResults int64
for _, r := range resp.reply.Responses {
numResults += r.GetInner().Header().NumKeys
}
if numResults > ba.MaxSpanRequestKeys {
panic(fmt.Sprintf("received %d results, limit was %d", numResults, ba.MaxSpanRequestKeys))
}
ba.MaxSpanRequestKeys -= numResults
// Exiting; fill in missing responses.
if ba.MaxSpanRequestKeys == 0 {
fillSkippedResponses(ba, resp.reply, seekKey)
return
}
}
}
// Check for completion.
if !ri.NeedAnother(rs) {
return
}
isFirst = false // next range will not be first!
rs = nextRS
}
// We've exited early. Return the range iterator error.
responseCh := make(chan response, 1)
responseCh <- response{pErr: ri.Error()}
responseChs = append(responseChs, responseCh)
return
}
// initAndVerifyBatch initializes timestamp-related information and
// verifies batch constraints before splitting.
func (ds *DistSender) initAndVerifyBatch(
ctx context.Context, ba *roachpb.BatchRequest,
) *roachpb.Error {
// In the event that timestamp isn't set and read consistency isn't
// required, set the timestamp using the local clock.
if ba.ReadConsistency == roachpb.INCONSISTENT && ba.Timestamp.Equal(hlc.ZeroTimestamp) {
ba.Timestamp = ds.clock.Now()
}
if ba.Txn != nil {
// Make a copy here since the code below modifies it in different places.
// TODO(tschottdorf): be smarter about this - no need to do it for
// requests that don't get split.
txnClone := ba.Txn.Clone()
ba.Txn = &txnClone
if len(ba.Txn.ObservedTimestamps) == 0 {
// Ensure the local NodeID is marked as free from clock offset;
// the transaction's timestamp was taken off the local clock.
if nDesc := ds.getNodeDescriptor(); nDesc != nil {
// TODO(tschottdorf): future refactoring should move this to txn
// creation in TxnCoordSender, which is currently unaware of the
// NodeID (and wraps *DistSender through client.Sender since it
// also needs test compatibility with *LocalSender).
//
// Taking care below to not modify any memory referenced from
// our BatchRequest which may be shared with others.
//
// We already have a clone of our txn (see above), so we can
// modify it freely.
//
// Zero the existing data. That makes sure that if we had
// something of size zero but with capacity, we don't re-use the
// existing space (which others may also use). This is just to
// satisfy paranoia/OCD and not expected to matter in practice.
ba.Txn.ResetObservedTimestamps()
// OrigTimestamp is the HLC timestamp at which the Txn started, so
// this effectively means no more uncertainty on this node.
ba.Txn.UpdateObservedTimestamp(nDesc.NodeID, ba.Txn.OrigTimestamp)
}
}
}
if len(ba.Requests) < 1 {
return roachpb.NewErrorf("empty batch")
}
if ba.MaxSpanRequestKeys != 0 {
// Verify that the batch contains only specific range requests or the
// Begin/EndTransactionRequest. Verify that a batch with a ReverseScan
// only contains ReverseScan range requests.
isReverse := ba.IsReverse()
for _, req := range ba.Requests {
inner := req.GetInner()
switch inner.(type) {
case *roachpb.ScanRequest, *roachpb.DeleteRangeRequest:
// Accepted range requests. All other range requests are still
// not supported.
// TODO(vivek): don't enumerate all range requests.
if isReverse {
return roachpb.NewErrorf("batch with limit contains both forward and reverse scans")
}
case *roachpb.BeginTransactionRequest, *roachpb.EndTransactionRequest, *roachpb.ReverseScanRequest:
continue
default:
return roachpb.NewErrorf("batch with limit contains %T request", inner)
}
}
}
return nil
}
// sendInternal sends the batch and updates the transaction on error. Depending
// on the error type, the transaction might be replaced by a new one.
func (txn *Txn) sendInternal(ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if len(ba.Requests) == 0 {
return nil, nil
}
if pErr := txn.db.prepareToSend(&ba); pErr != nil {
return nil, pErr
}
// Send call through the DB's sender.
ba.Txn = &txn.Proto
// For testing purposes, txn.UserPriority can be a negative value (see
// MakePriority).
if txn.UserPriority != 0 {
ba.UserPriority = txn.UserPriority
}
// TODO(radu): when db.send supports a context, we can just use that (and
// remove the prepareToSend call above).
br, pErr := txn.db.sender.Send(txn.Context, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(txn.db.sender, br))
}
if br != nil {
for _, encSp := range br.CollectedSpans {
var newSp basictracer.RawSpan
if err := tracing.DecodeRawSpan(encSp, &newSp); err != nil {
return nil, roachpb.NewError(err)
}
txn.CollectedSpans = append(txn.CollectedSpans, newSp)
}
}
// Only successful requests can carry an updated Txn in their response
// header. Any error (e.g. a restart) can have a Txn attached to them as
// well; those update our local state in the same way for the next attempt.
// The exception is if our transaction was aborted and needs to restart
// from scratch, in which case we do just that.
if pErr == nil {
txn.Proto.Update(br.Txn)
return br, nil
}
if log.V(1) {
log.Infof(txn.Context, "failed batch: %s", pErr)
}
if _, ok := pErr.GetDetail().(*roachpb.TransactionAbortedError); ok {
// On Abort, reset the transaction so we start anew on restart.
txn.Proto = roachpb.Transaction{
TxnMeta: enginepb.TxnMeta{
Isolation: txn.Proto.Isolation,
},
Name: txn.Proto.Name,
}
// Acts as a minimum priority on restart.
if pErr.GetTxn() != nil {
txn.Proto.Priority = pErr.GetTxn().Priority
}
} else if pErr.TransactionRestart != roachpb.TransactionRestart_NONE {
txn.Proto.Update(pErr.GetTxn())
}
return nil, pErr
}
func (tc *TxnCoordSender) heartbeat(ctx context.Context, txnID uuid.UUID) bool {
tc.Lock()
txnMeta := tc.txns[txnID]
txn := txnMeta.txn.Clone()
hasAbandoned := txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow())
tc.Unlock()
if txn.Status != roachpb.PENDING {
// A previous iteration has already determined that the transaction is
// already finalized, so we wait for the client to realize that and
// want to keep our state for the time being (to dish out the right
// error once it returns).
return true
}
// Before we send a heartbeat, determine whether this transaction should be
// considered abandoned. If so, exit heartbeat. If ctx.Done() is not nil, then
// it is a cancellable Context and we skip this check and use the ctx lifetime
// instead of a timeout.
//
// TODO(andrei): We should disallow non-cancellable contexts in the heartbeat
// goroutine and enforce that our kv client cancels the context when it's
// done. We get non-cancellable contexts from remote clients
// (roachpb.ExternalClient) because we override the gRPC context to make it
// non-cancellable in DBServer.Batch (as that context is not tied to a txn
// lifetime).
// Further note that, unfortunately, the Sender interface generally makes it
// difficult for the TxnCoordSender to get a context with the same lifetime as
// the transaction (the TxnCoordSender associates the context of the txn's
// first write with the txn). We should move to using only use local clients
// (i.e. merge, or at least co-locate client.Txn and the TxnCoordSender). At
// that point, we probably don't even need to deal with context cancellation
// any more; the client will be trusted to always send an EndRequest when it's
// done with a transaction.
if ctx.Done() == nil && hasAbandoned {
if log.V(1) {
log.Infof(ctx, "transaction %s abandoned; stopping heartbeat", txnMeta.txn)
}
tc.tryAsyncAbort(txnID)
return false
}
ba := roachpb.BatchRequest{}
ba.Txn = &txn
hb := &roachpb.HeartbeatTxnRequest{
Now: tc.clock.Now(),
}
hb.Key = txn.Key
ba.Add(hb)
log.Event(ctx, "heartbeat")
br, pErr := tc.wrapped.Send(ctx, ba)
// Correctness mandates that when we can't heartbeat the transaction, we
// make sure the client doesn't keep going. This is particularly relevant
// in the case of an ABORTED transaction, but if we can't reach the
// transaction record at all, we're going to have to assume we're aborted
// as well.
if pErr != nil {
log.Warningf(ctx, "heartbeat to %s failed: %s", txn, pErr)
// We're not going to let the client carry out additional requests, so
// try to clean up.
tc.tryAsyncAbort(*txn.ID)
txn.Status = roachpb.ABORTED
} else {
txn.Update(br.Responses[0].GetInner().(*roachpb.HeartbeatTxnResponse).Txn)
}
// Give the news to the txn in the txns map. This will update long-running
// transactions (which may find out that they have to restart in that way),
// but in particular makes sure that they notice when they've been aborted
// (in which case we'll give them an error on their next request).
tc.Lock()
tc.txns[txnID].txn.Update(&txn)
tc.Unlock()
return true
}
func (tc *TxnCoordSender) heartbeat(ctx context.Context, txnID uuid.UUID) bool {
tc.Lock()
txnMeta := tc.txns[txnID]
txn := txnMeta.txn.Clone()
hasAbandoned := txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow())
tc.Unlock()
if txn.Status != roachpb.PENDING {
// A previous iteration has already determined that the transaction is
// already finalized, so we wait for the client to realize that and
// want to keep our state for the time being (to dish out the right
// error once it returns).
return true
}
// Before we send a heartbeat, determine whether this transaction should be
// considered abandoned. If so, exit heartbeat. If ctx.Done() is not nil, then
// it is a cancelable Context and we skip this check and use the ctx lifetime
// instead of a timeout.
if ctx.Done() == nil && hasAbandoned {
if log.V(1) {
log.Infof(ctx, "transaction %s abandoned; stopping heartbeat", txnMeta.txn)
}
tc.tryAsyncAbort(txnID)
return false
}
ba := roachpb.BatchRequest{}
ba.Txn = &txn
hb := &roachpb.HeartbeatTxnRequest{
Now: tc.clock.Now(),
}
hb.Key = txn.Key
ba.Add(hb)
log.Event(ctx, "heartbeat")
br, pErr := tc.wrapped.Send(ctx, ba)
// Correctness mandates that when we can't heartbeat the transaction, we
// make sure the client doesn't keep going. This is particularly relevant
// in the case of an ABORTED transaction, but if we can't reach the
// transaction record at all, we're going to have to assume we're aborted
// as well.
if pErr != nil {
log.Warningf(ctx, "heartbeat to %s failed: %s", txn, pErr)
// We're not going to let the client carry out additional requests, so
// try to clean up.
tc.tryAsyncAbort(*txn.ID)
txn.Status = roachpb.ABORTED
} else {
txn.Update(br.Responses[0].GetInner().(*roachpb.HeartbeatTxnResponse).Txn)
}
// Give the news to the txn in the txns map. This will update long-running
// transactions (which may find out that they have to restart in that way),
// but in particular makes sure that they notice when they've been aborted
// (in which case we'll give them an error on their next request).
tc.Lock()
tc.txns[txnID].txn.Update(&txn)
tc.Unlock()
return true
}
