Buggy sink returning a checkpoint that is too large can cause flow control to dead-lock

[liuzix]

What did you do?

When a table pipeline is blocked by flow control when waiting for the global checkpoint to reach a DDL commit-ts, a table moving into the processor where the blocked table pipeline is running can cause dead lock. If the sink is buggy so that it returns a checkpoint that is greater than the argument of FlushRowChangedEvents.

What did you expect to see?

We do not expect a buggy sink. But even if the sink were buggy, we would not expect to see a deadlock problem.

What did you see instead?

The table pipeline is deadlocked. And the DDL is executed prematurely.

Versions of the cluster

TiCDC version (execute cdc version):

f847b331572379527bf37a7f19be20448a74b2c2 on release-5.3

[overvenus]

Inaccurate checkpoint bug has been fixed.