Hi Robson, Thank you for the detailed bug report.
I believe the behavior that you're describing is caused by this flaw: https://issues.apache.org/jira/browse/KAFKA-15090 which is still under discussion. Since the above flaw was introduced in 3.0, source connectors need to return from poll() before the graceful shutdown timeout to avoid the error. You may be able to mitigate the error if the connector allows you to reduce its poll timeout/interval to something less than the graceful timeout, but that will depend on the specific connector implementation, so check the documentation for your connector. I know some implementations have received patches to compensate for this behavior in the framework, so also consider upgrading or checking release notes for your connectors. As for the effects of this error: whenever a non-graceful stop occurs, the runtime will immediately close the producer so that the task will not be able to write any further records. However, it will still leave resources for that task (threads, in-memory records, database connections, etc) allocated, until the task does finally return from poll(). While this is not desirable behavior, it seems to be treated as just a nuisance error by most operators. I hope this gives some context for the error message you're seeing. Thanks, Greg