StefanRRichter edited a comment on issue #7009: [FLINK-10712] Support to restore state when using RestartPipelinedRegionStrategy URL: https://github.com/apache/flink/pull/7009#issuecomment-452649422 I have one more concern that might lead to bugs in a certain corner case. What will happen in your change if the task is using operator state, union state in particular. In `applyRepartitioner()`, https://github.com/apache/flink/blob/1e2aa8e9f35e7a943a4ed56a47834ee50bab3b47/flink-runtime/src/main/java/org/apache/flink/runtime/checkpoint/StateAssignmentOperation.java#L636 you can see that all operator state is repartitioned if there is one union state. this can already be a problem with a union state, but even more if there is a union state and some partionable state - the partitioning for the partitionable state for those tasks that are restarted could differ to the partitioning used in the original run - some partitions could be dropped or assigned twice by this. I think that means we need to change the method to only redistribute the union states, and I wonder if even distributioing the union state only for the failed task even makes sense. I think we need a testcase for this scenario (operator with 1 union and 1 partionable operator state)and I think it might fail as described before when we check how the operator state was reassigned after some partial recovery. What do you think?
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
