Github user StephanEwen commented on the issue: https://github.com/apache/flink/pull/1668 To suggest some way to fix the guarantees: To my mind, the crux lies in the way that the feedback channel is implemented - a simple blocking queue just does not cut it for that case. To make this proper, I think we need to do the following: - Have an elastic feedback channel (unbounded) with a certain memory budget, that can spill if needed. I think it would be best implemented holding data serialized. - On checkpoint, one simply adds the feedback channel data (already bytes) to the checkpoint - The source task should probably prioritize reading from the feedback channel, to keep it always as small as possible.
--- 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 infrastruct...@apache.org or file a JIRA ticket with INFRA. ---