[jira] [Commented] (KAFKA-3470) Consumer group coordinator should take commit requests as effective as heartbeats

["Jason Gustafson (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-3470?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15232663#comment-15232663 ] 

Jason Gustafson commented on KAFKA-3470:
----------------------------------------

I think this makes sense. In some situations that I've seen in testing with the Java client, when replication of offset commits is slow (e.g. because of a broker failure), you can get a couple commits effectively blocking out heartbeats. Eventually when the heartbeat gets through, you may find that the group has already begun rebalancing. So treating commits as heartbeats closes the window a little bit for such failures. It also seems intuitive to treat commits as an indication of consumer liveness. I'll go ahead and submit a patch today.



> Consumer group coordinator should take commit requests as effective as heartbeats
> ---------------------------------------------------------------------------------
>
>                 Key: KAFKA-3470
>                 URL: https://issues.apache.org/jira/browse/KAFKA-3470
>             Project: Kafka
>          Issue Type: Improvement
>            Reporter: Zaiming Shi
>
> Group coordinator does not reset heartbeat timer when commit request is received.
> This may lead to unnecessary session timeouts when a consumer sends
> sync commit requests so frequently that causes heartbeat requests to be
> delayed.
> Presumably (as I do not know Kafka code well):
> Commit requests (v1 and v2) have all data fields for a heartbeat request,
> If they are taken as effective as heartbeat requests, we should have
> better group stability.
> [For reference]
> previous discussions in: users@kafka.apache.org
> mail title: 	consumer group, why commit requests are not considered as effective heartbeats?



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)