[
https://issues.apache.org/jira/browse/KAFKA-13217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17402859#comment-17402859
]
Guozhang Wang commented on KAFKA-13217:
---------------------------------------
Thanks [~ableegoldman] I think this is a good time to discuss about it: the
concern I had about sending leave group immediately is not about the
rebalance's assignor overhead itself, but that the first rebalance would move
the tasks on that shutdown instance to others immediately, which would start
restoring the states --- even if we have standby tasks which could be promote
as active, we still need a new standby replica and whoever gets it would need
to do restoration. With the current optimization, we would not migrate tasks,
and when the "second" (we actually only trigger rebalance once right now, but
just for presentation) rebalance would likely gets the tasks back to the
restarted instances and hence no task migration at all.
That being said, I like the idea of separating permanent closure v.s. bouncing
with separate APIs as it seems to be the right behavior.
> Reconsider skipping the LeaveGroup on close() or add an overload that does so
> -----------------------------------------------------------------------------
>
> Key: KAFKA-13217
> URL: https://issues.apache.org/jira/browse/KAFKA-13217
> Project: Kafka
> Issue Type: Improvement
> Components: streams
> Reporter: A. Sophie Blee-Goldman
> Priority: Major
>
> In Kafka Streams, when an instance is shut down via the close() API, we
> intentionally skip sending a LeaveGroup request. This is because often the
> shutdown is not due to a scaling down event but instead some transient
> closure, such as during a rolling bounce. In cases where the instance is
> expected to start up again shortly after, we originally wanted to avoid that
> member's tasks from being redistributed across the remaining group members
> since this would disturb the stable assignment and could cause unnecessary
> state migration and restoration. We also hoped
> to limit the disruption to just a single rebalance, rather than forcing the
> group to rebalance once when the member shuts down and then again when it
> comes back up. So it's really an optimization for the case in which the
> shutdown is temporary.
>
> That said, many of those optimizations are no longer necessary or at least
> much less useful given recent features and improvements. For example
> rebalances are now lightweight so skipping the 2nd rebalance is not as worth
> optimizing for, and the new assignor will take into account the actual
> underlying state for each task/partition assignment, rather than just the
> previous assignment, so the assignment should be considerably more stable
> across bounces and rolling restarts.
>
> Given that, it might be time to reconsider this optimization. Alternatively,
> we could introduce another form of the close() API that forces the member to
> leave the group, to be used in event of actual scale down rather than a
> transient bounce.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
