// 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(0)
clock := hlc.NewClock(manual.UnixNano)
clock.SetMaxOffset(20)
ts := NewTxnCoordSender(senderFn(func(_ context.Context, ba proto.BatchRequest) (*proto.BatchResponse, *proto.Error) {
return ba.CreateReply().(*proto.BatchResponse), nil
}), clock, false, nil, stopper)
// 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 proto.BatchRequest
put := &proto.PutRequest{}
put.Key = proto.Key("test")
ba.Add(put)
ba.Add(&proto.EndTransactionRequest{})
ba.Txn = &proto.Transaction{Name: "test"}
_, pErr := ts.Send(context.Background(), ba)
if pErr != nil {
t.Fatal(pErr)
}
}
func (ts *txnSender) Send(ctx context.Context, ba proto.BatchRequest) (*proto.BatchResponse, *proto.Error) {
// Send call through wrapped sender.
ba.Txn = &ts.Proto
br, pErr := ts.wrapped.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(proto.ErrorUnexpectedlySet(ts.wrapped, br))
}
// TODO(tschottdorf): see about using only the top-level *proto.Error
// information for this restart logic (includes adding the Txn).
err := pErr.GoError()
// 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 err == nil {
ts.Proto.Update(br.Txn)
return br, nil
} else if abrtErr, ok := err.(*proto.TransactionAbortedError); ok {
// On Abort, reset the transaction so we start anew on restart.
ts.Proto = proto.Transaction{
Name: ts.Proto.Name,
Isolation: ts.Proto.Isolation,
// Acts as a minimum priority on restart.
Priority: abrtErr.Transaction().GetPriority(),
}
} else if txnErr, ok := err.(proto.TransactionRestartError); ok {
ts.Proto.Update(txnErr.Transaction())
}
return nil, pErr
}
func (tc *TxnCoordSender) heartbeat(id string, trace *tracer.Trace, ctx context.Context) bool {
tc.Lock()
proceed := true
txnMeta := tc.txns[id]
// Before we send a heartbeat, determine whether this transaction
// should be considered abandoned. If so, exit heartbeat.
if txnMeta.hasClientAbandonedCoord(tc.clock.PhysicalNow()) {
// TODO(tschottdorf): should we be more proactive here?
// The client might be continuing the transaction
// through another coordinator, but in the most likely
// case it's just gone and the open transaction record
// could block concurrent operations.
if log.V(1) {
log.Infof("transaction %s abandoned; stopping heartbeat",
txnMeta.txn)
}
proceed = false
}
// txnMeta.txn is possibly replaced concurrently,
// so grab a copy before unlocking.
txn := txnMeta.txn
tc.Unlock()
if !proceed {
return false
}
hb := &proto.HeartbeatTxnRequest{}
hb.Key = txn.Key
ba := proto.BatchRequest{}
ba.Timestamp = tc.clock.Now()
ba.Key = txn.Key
ba.Txn = &txn
ba.Add(hb)
epochEnds := trace.Epoch("heartbeat")
_, err := tc.wrapped.Send(ctx, ba)
epochEnds()
// If the transaction is not in pending state, then we can stop
// the heartbeat. It's either aborted or committed, and we resolve
// write intents accordingly.
if err != nil {
log.Warningf("heartbeat to %s failed: %s", txn, err)
}
// TODO(bdarnell): once we have gotten a heartbeat response with
// Status != PENDING, future heartbeats are useless. However, we
// need to continue the heartbeatLoop until the client either
// commits or abandons the transaction. We could save a little
// pointless work by restructuring this loop to stop sending
// heartbeats between the time that the transaction is aborted and
// the client finds out. Furthermore, we could use this information
// to send TransactionAbortedErrors to the client so it can restart
// immediately instead of running until its EndTransaction.
return true
}
// 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.
func (tc *TxnCoordSender) maybeBeginTxn(ba *proto.BatchRequest) {
if ba.Txn == nil {
return
}
if len(ba.Requests) == 0 {
panic("empty batch with txn")
}
if len(ba.Txn.ID) == 0 {
// TODO(tschottdorf): should really choose the first txn write here.
firstKey := ba.Requests[0].GetInner().Header().Key
newTxn := proto.NewTransaction(ba.Txn.Name, keys.KeyAddress(firstKey), ba.GetUserPriority(),
ba.Txn.Isolation, tc.clock.Now(), 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
}
}
