Page History

...

The reason for doing two rolling bounces is because the old transactional producer doesn’t have access to consumer generation, so group coordinator doesn’t have an effective way to fence old zombies. By doing first rolling bounce, the task producer will also opt in accessing the consumer state and send TxnOffsetCommitRequest with generation. With this foundational change, it is much safer to execute step 3.

***Note that the above recommended upgrade path is for users who need consistency guarantee. For users who don't worry about consistency, step 2 & 3 could be combined into a single rolling bounce with Kafka client library upgrade. The application should resume work without problem.

Non-stream EOS  Upgrade

As for non-stream users, they would require following steps:

...