Github user gyfora commented on the pull request:
https://github.com/apache/flink/pull/937#issuecomment-125125052
After playing around with the StreamCheckpointingITCase, it seems like we
actually don't even have at least once guarantees because of some bug in the
implementation (either in the barrier buffer or with the event alignment):
The RichFlatMap that did the counting (preceeded by a groupBy) did not
always count the correct number of inputs or prefixes (meaning the state was
checkpointed earlier than it should have been).
What was interesting that all other operators counted exactly the right
amount of inputs all the time, then I realised they were all forward connected
(pointwise connection pattern which didnt trigger any blocking logic in the
barrier buffer). So I changed the connection between the filter and map to
shuffle, and now that map also fails sometimes on incorrect number of inputs
received.
I will try rebasing this on @StephanEwen 's barrier buffer rework in
https://github.com/apache/flink/pull/938, let's see if that fixes it.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
