[ https://issues.apache.org/jira/browse/KAFKA-21380 ]
Sushant Mahajan deleted comment on KAFKA-21380:
-----------------------------------------
was (Author: JIRAUSER300816):
This was deliberately avoided as a persister fail => a 5 time call to persister
with exponential backoff failed. Now with DLQ again, the first step is
ARCHIVING which requires a persister write. This too has good chance of failing
(same partition key, so same broker). Hence the decision was taken to not stall
the in-flight records queue over this.
> Failed state-transition rollback bypasses DLQ when delivery count limit is
> reached
> ----------------------------------------------------------------------------------
>
> Key: KAFKA-21380
> URL: https://issues.apache.org/jira/browse/KAFKA-21380
> Project: Kafka
> Issue Type: Bug
> Components: consumer
> Affects Versions: 4.5.0
> Reporter: da Yi
> Assignee: da Yi
> Priority: Major
>
> When the DLQ flow is enabled for a share group (KIP-1191), records whose
> delivery
> count reaches the limit are moved to ARCHIVING on the normal release path so
> they
> are written to the DLQ topic (InFlightState.tryUpdateState with
> dlqSupportEnabled=true).
> However, the rollback path in InFlightState.completeStateTransition(false) -
> taken
> when a two-phase state transition fails (e.g. the persister write fails) and
> the
> acquisition lock has expired - archives the records directly:
> state = previousState.deliveryCount() >= rollbackState.maxDeliveryCount ?
> RecordState.ARCHIVED : RecordState.AVAILABLE;
> This bypasses the DLQ flow entirely: records that should have been
> dead-lettered are
> silently archived and never appear in the DLQ topic. Two related
> inconsistencies:
> 1. ARCHIVING (normal path) vs ARCHIVED (rollback path) for the same
> limit-reached
> condition.
> 2. The normal path evaluates the limit with the delivery count including the
> current
> delivery, while the rollback path uses previousState.deliveryCount() - the
> counts
> differ by one.
> Note that completeStateTransition(boolean) has no access to the
> dlqSupportEnabled
> flag, so the rollback path cannot currently reach ARCHIVING.
> Reproduction: new unit test InFlightStateTest
> (testFailedStateTransitionWithExpiredLockDeadLettersWhenDlqSupported) fails on
> trunk and passes with the fix.
> Proposed fix: thread the DLQ-support flag through the rollback path,
> mirroring the
> normal transition path, and start the DLQ flow for batches that move to
> ARCHIVING
> in both the local rollback and the persister-failure paths.
> Two test scenarios (step 5 in the reproduction) verify:
> - normal release path at the limit moves records to ARCHIVING (baseline)
> - rollback with expired lock at the limit moves records to ARCHIVING when DLQ
> is
> enabled, and to ARCHIVED when disabled
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
