[
https://issues.apache.org/jira/browse/KAFKA-20719?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18090384#comment-18090384
]
Matthias J. Sax commented on KAFKA-20719:
-----------------------------------------
I opened a PR with a new integration test to demonstrate the issue:
[https://github.com/apache/kafka/pull/22630]
> "Streams" group sticky assignor not sticky at startup
> -----------------------------------------------------
>
> Key: KAFKA-20719
> URL: https://issues.apache.org/jira/browse/KAFKA-20719
> Project: Kafka
> Issue Type: Bug
> Components: group-coordinator, streams
> Affects Versions: 4.2.0
> Reporter: Matthias J. Sax
> Priority: Critical
>
> For "streams" groups (KIP-1071), the sticky-assignor is not sticky at startup
> of a new KS instance/client.
> In the "classic" protocol, a newly starting client encodes what task
> directories it finds in its local state directory as part of the rebalance
> metadata. This allows the "classic" client-side sticky assignor to re-assign
> tasks to previous clients, meeting the goal of being sticky.
> In contrast, for "streams" group no such information is encoded. A starting
> instance correctly sends empty previous assignment on its initial "join group
> heartbeat" (with epoch=0). While the group coordinator does pass in the
> "current/previous assignment" into the sticky assignor, new clients/threads
> get a new memberId, and thus previous state cannot be re-assigned in a sticky
> manner on startup. (Note, static membership might address this, but it's also
> not yet supported for "streams" group:
> https://issues.apache.org/jira/browse/KAFKA-20169)
> To fix this issue, we should encode offset information(derived from the local
> task offset checkpoint) via `taskOffsetSum` field in the "join group
> heartbeat", and update the sticky-assignor to use this data.
> This ticket is blocked by https://issues.apache.org/jira/browse/KAFKA-20116
> which does add support for `taskOffsetSum` client and broker side already, to
> support warmup tasks. After KAFKA-20116 was completed, this ticket would need
> to fix the sticky-assignor itself to use the newly available data.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
