source-postgres: Fix watchdog timeout error reporting

The intent of this logic was that if and only if the replication
streaming is interrupted by the "we really should have reached the
established fence by now" watchdog timeout, we ought to report an
appropriate error. But I was checking the cancellation cause on
the wrong context, so that wouldn't have worked as intended, and
also it turns out that `cancel(nil)` sets the cause to `Canceled`
so a bit more logic is required to turn that into a nil error
return from the streaming operation anyway.

Author: willdonnelly

source-postgres/replication.go

@@ -348,6 +348,7 @@ func (s *replicationStream) StreamToFence(ctx context.Context, fenceAfter time.D
 
 	// Stream replication events until the fence is reached or the watchdog timeout hits.
 	var relayCtx, cancelRelayCtx = context.WithCancelCause(ctx)
+	var fenceReached = errors.New("fenced reached")
 	defer cancelRelayCtx(nil)
 
 	// Given that the early-exit fast path was not taken, there must be further data for
@@ -387,12 +388,15 @@ func (s *replicationStream) StreamToFence(ctx context.Context, fenceAfter time.D
 			if eventLSN >= fenceLSN {
 				logrus.WithField("cursor", eventLSN.String()).Debug("finished fenced streaming phase")
 				s.previousFenceLSN = eventLSN
-				cancelRelayCtx(nil) // Stop the relay loop so we can exit cleanly. A nil cause means no error.
+				cancelRelayCtx(fenceReached) // Stop the relay loop so we can exit cleanly.
 			}
 		}
 		return nil
 	}); errors.Is(err, context.Canceled) {
-		return context.Cause(ctx)
+		if cause := context.Cause(relayCtx); !errors.Is(cause, fenceReached) {
+			return cause
+		}
+		return nil
 	} else if err != nil {
 		return err
 	}