[
https://issues.apache.org/jira/browse/KAFKA-12462?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17300654#comment-17300654
]
A. Sophie Blee-Goldman commented on KAFKA-12462:
------------------------------------------------
The only real downside in 2.7 is that we won't properly clean up the task, ie
we'll skip committing the offsets and writing the checkpoint. So we'd lose any
work we did since the last commit – for EOS this would be a perf hit since we'd
probably need to restore the state stores from scratch after starting back up,
whereas for ALOS we could get some overcounting. Not the end of the world, but
worth fixing if we can
> Threads in PENDING_SHUTDOWN entering a rebalance can cause an illegal state
> exception
> --------------------------------------------------------------------------------------
>
> Key: KAFKA-12462
> URL: https://issues.apache.org/jira/browse/KAFKA-12462
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 2.8.0
> Reporter: Walker Carlson
> Assignee: A. Sophie Blee-Goldman
> Priority: Blocker
> Labels: streams
> Fix For: 2.8.0, 2.7.1
>
>
> A thread was removed, sending it to the PENDING_SHUTDOWN state, but went
> through a rebalance before completing the shutdown.
> {code:java}
> // [2021-03-07 04:33:39,385] DEBUG [i-07430efc31ad166b7-StreamThread-6]
> stream-thread [i-07430efc31ad166b7-StreamThread-6] Ignoring request to
> transit from PENDING_SHUTDOWN to PARTITIONS_REVOKED: only DEAD state is a
> valid next state (org.apache.kafka.streams.processor.internals.StreamThread)
> {code}
> Inside StreamsRebalanceListener#onPartitionsRevoked, we have
> {code:java}
> //
> if (streamThread.setState(State.PARTITIONS_REVOKED) != null &&
> !partitions.isEmpty())
> taskManager.handleRevocation(partitions);
> {code}
> Since PENDING_SHUTDOWN → PARTITIONS_REVOKED is a disallowed transition, we
> never invoke TaskManager#handleRevocation. Currently handleRevocation is
> responsible for preparing any active tasks for close, including committing
> offsets and writing the checkpoint as well as suspending the task. We can’t
> close the task in handleRevocation since we still support EAGER rebalancing,
> which invokes handleRevocation at the beginning of a rebalance on all tasks.
> The tasks that are actually revoked will be closed during
> TaskManager#handleAssignment . The IllegalStateException is specifically
> because we don’t suspend the task before attempting to close it, and the
> direct transition from RUNNING → CLOSED is forbidden.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
