Open kamicut opened 6 years ago
Ideally it can get through at least a minute during each invocation, so the checkpoint should be able to progress, albeit slowly. (I would rather re-publish changes for "at least once delivery" instead of dropping them.)
One snag though is that the checkpointing occurs after the source data has been push
ed onto the stream, which isn't a guarantee that it was consumed (and published), so maybe we need to think about that too.
What should the behavior in the following scenario:
Should the data be declared lost and the checkpoint reset?
cc @mojodna