Flink marks source late arriving events

38 Views Asked by At

We're building a Flink app that consumes events from different Kafka topics. This app uses the bounded out of order watermark strategy on the source. During normal execution everything works as expected and we do not get any late arriving data (based on watermarks), but on checkpoints/ savepoints restores we're getting late arriving events, no matter how much we increase the out of order bound. Did anyone ever encounter this situation?

1

There are 1 best solutions below

8
David Anderson On

Watermarks are not checkpointed, so after recovery, the watermarks have to be re-established based on the events processed after the checkpoint. If one or more sources are more-or-less idle at that point in time, this could explain why the behavior is so different compared to the situation before the restart.