David Jacot created KAFKA-17507:
-----------------------------------
Summary: WriteTxnMarkers API must not return until markers are
written and materialized in group coordinator's cache
Key: KAFKA-17507
URL: https://issues.apache.org/jira/browse/KAFKA-17507
Project: Kafka
Issue Type: Bug
Reporter: David Jacot
Assignee: David Jacot
We have observed the below errors in some cluster:
Uncaught exception in scheduled task 'handleTxnCompletion-902667'
exception.message:Trying to complete a transactional offset commit for
producerId *** and groupId *** even though the offset commit record itself
hasn't been appended to the log.
When a transaction is completed, the transaction coordinator sends a
WriteTxnMarkers request to all the partitions involved in the transaction to
write the markers to them. When the broker receives it, it writes the markers
and if markers are written to the __consumer_offsets partitions, it informs the
group coordinator that it can materialize the pending transactional offsets in
its main cache. The group coordinator does this asynchronously since Apache
Kafka 2.0, see this
[patch|https://github.com/apache/kafka/commit/c53e274d3128bc92f0e8b6a79c407cf764f16f7b].
The above error appends when the asynchronous operation is executed by the
scheduler and the operation finds that there are pending transactional offsets
that were not written yet. How come?
There is actually an issue is the steps described above. The group coordinator
does not wait until the asynchronous operation completes to return to the api
layer. Hence the WriteTxnMarkers response may be send back to the transaction
coordinator before the async operation is actually completed. Hence it is
possible that the next transactional produce to be started also before the
operation is completed too. This could explain why the group coordinator has
pending transactional offsets that are not written yet.
There is a similar issue when the transaction is aborted. However on this path,
we don't have any checks to verify whether all the pending transactional
offsets have been written or not so we don't see any errors in our logs. Due to
the same race condition, it is possible to actually remove the wrong pending
transactional offsets.
PS: The new group coordinator is not impacted by this bug.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
