Anjaneya Gupta created KAFKA-20914:
--------------------------------------
Summary: KafkaRaftClient does not clear append/fetch purgatories
when a leader transitions to Resigned or Unattached
Key: KAFKA-20914
URL: https://issues.apache.org/jira/browse/KAFKA-20914
Project: Kafka
Issue Type: Improvement
Reporter: Anjaneya Gupta
When a node stops being the leader, its parked requests can no longer be
completed normally. It can't commit appends anymore, and held fetches should be
failed so callers retry against the new leader. So both purgatories should be
completed exceptionally when a leader loses leadership. onBecomeFollower does
this, but the other two leader-exit paths do not i.e transitionToResigned does
not clear it in appendPurgatory and transitionToUnattached does not clear it in
both fetchPurgatory and appendPurgatory.
Impact: The pending futures eventually expire via request.timeout.ms. But until
then, callers wait the full timeout for a failure that is already certain, and
each pending append future keeps its CompletedBatch (and its backing
ByteBuffer) in memory the whole time.
Fix: complete both purgatories exceptionally on every transition from leader to
non-leader.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
