dawidwys opened a new pull request #16089: URL: https://github.com/apache/flink/pull/16089
This is a backport of parts of #16019 and #15294 Broadcast partitioning can not work with unaligned checkpointing. There are no guarantees that records are consumed at the same rate in all channels. This can result in some tasks applying state changes corresponding to a certain broadcasted event while others don't. In turn upon restore, it may lead to an inconsistent state. This is especially harmful to the most common pattern for using a broadcast pattern where only a single operator checkpoints its state and that state is copied over to all other operators on restore. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
