Hi.
I’ve run into a problem with Kafka consumers not leaving their consumer group cleanly when the application restarts. Out of 9 topics that (with a consumer for each) it seems that every time at least 2 or 3 do not leave the group cleanly so when the application starts these consumers start consuming only after session.timeout.ms expires and the group re-balances. When debugging this I’ve enabled logging from GroupCoordinator in the broker and I’ve really observed that not all of LeaveGroupRequests were received. I’ve looked at the code in KafkaConsumer.close() and the effort to send the LeaveGroupRequest really is minimal (and the comments suggest this is the intention). So if the network client cannot send the request immediately, it is not sent at all. But I think that the amount of effort put into sending LeaveGroupRequest should be the decision of the consumer’s user, not the library itself. IMHO it would be enough if the poll timeout when sending LeaveGroupRequest was configurable. We’re using Kafka 0.10.0.1. I’ve looked at the KafkaConsumer.close() code in the master branch and it looks like is has changed quite a bit when compared to 0.10.0.1 but the LeaveGroupRequest sending effort has not changed IMHO. Vlastimil
