Hi! Flink

1.17.1

, using SQL, we have issues completing checkpoints. We’ve tried enabled unaligned and not, and its not an issue with timing out, the checkpointing failure happens before the specified checkpoint timeout. The weird/interesting thing is, even with unaligned checkpoints, we have several operators (JOIN) that are back-pressured but never even acknowledge the checkpoint… Has anyone witnessed anything similar? thnx!

If the operators are back pressured, what is the operator after that one? Because that’s most likely the bottleneck

The final Sink has 100% completed checkpoints, but some of the the JOINS preceding don’t even ack…

I think they are under so much back pressure because of the final sink, they can’t ack. Not sure what you’re materializing, but that can be the culprit

👍 thnx! that is what we were wanting to confirm

Are you using the upsert-Kafka connector?

yes

One thing worth considering is buffering records before flushing them to Kafka, to avoid that you need to write multiple updates for the same record

This occurs during backfill/snapshot read