func (tc *TxnCoordSender) heartbeat(txnID uuid.UUID, trace opentracing.Span, ctx context.Context) bool {
tc.Lock()
proceed := true
txnMeta := tc.txns[txnID.String()]
// 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 := &roachpb.HeartbeatTxnRequest{}
hb.Key = txn.Key
ba := roachpb.BatchRequest{}
ba.Timestamp = tc.clock.Now()
txnClone := txn.Clone()
ba.Txn = &txnClone
ba.Add(hb)
trace.LogEvent("heartbeat")
_, err := tc.wrapped.Send(ctx, ba)
// 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
}
// heartbeatLoop periodically sends a HeartbeatTxn RPC to an extant
// transaction, stopping in the event the transaction is aborted or
// committed after attempting to resolve the intents. When the
// heartbeat stops, the transaction is unregistered from the
// coordinator,
func (tc *TxnCoordSender) heartbeatLoop(txnID uuid.UUID) {
var tickChan <-chan time.Time
{
ticker := time.NewTicker(tc.heartbeatInterval)
tickChan = ticker.C
defer ticker.Stop()
}
defer func() {
tc.Lock()
tc.unregisterTxnLocked(txnID)
tc.Unlock()
}()
var closer <-chan struct{}
var sp opentracing.Span
{
tc.Lock()
txnMeta := tc.txns[txnID.String()] // do not leak to outer scope
closer = txnMeta.txnEnd
sp = tc.tracer.StartTrace("heartbeat loop")
defer sp.Finish()
tc.Unlock()
}
if closer == nil {
// Avoid race in which a Txn is cleaned up before the heartbeat
// goroutine gets a chance to start.
return
}
ctx, _ := opentracing.ContextWithSpan(context.Background(), sp)
// Loop with ticker for periodic heartbeats.
for {
select {
case <-tickChan:
if !tc.heartbeat(txnID, sp, ctx) {
return
}
case <-closer:
// Transaction finished normally.
return
case <-tc.stopper.ShouldDrain():
return
}
}
}
// unregisterTxn deletes a txnMetadata object from the sender
// and collects its stats. It assumes the lock is held.
func (tc *TxnCoordSender) unregisterTxnLocked(txnID uuid.UUID) {
txnIDStr := txnID.String()
txnMeta := tc.txns[txnIDStr] // guaranteed to exist
if txnMeta == nil {
panic(fmt.Sprintf("attempt to unregister non-existent transaction: %s", txnID))
}
tc.txnStats.durations = append(tc.txnStats.durations, float64(tc.clock.PhysicalNow()-txnMeta.firstUpdateNanos))
tc.txnStats.restarts = append(tc.txnStats.restarts, float64(txnMeta.txn.Epoch))
switch txnMeta.txn.Status {
case roachpb.ABORTED:
tc.txnStats.aborted++
case roachpb.PENDING:
tc.txnStats.abandoned++
case roachpb.COMMITTED:
tc.txnStats.committed++
}
txnMeta.keys.Clear()
delete(tc.txns, txnIDStr)
}