[
https://issues.apache.org/jira/browse/KAFKA-21516?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
hechengbo updated KAFKA-21516:
------------------------------
Attachment: kafka-share-initialize-jira-reproducer.patch
> Delayed InitializeShareGroupState retries can overwrite persisted consumption
> progress
> --------------------------------------------------------------------------------------
>
> Key: KAFKA-21516
> URL: https://issues.apache.org/jira/browse/KAFKA-21516
> Project: Kafka
> Issue Type: Bug
> Affects Versions: 4.5.0
> Reporter: hechengbo
> Priority: Major
> Attachments: kafka-share-initialize-jira-reproducer.patch
>
>
> h1. Problem
> ShareCoordinatorShard.initializeState treats an initialization request as a
> duplicate only when both stateEpoch and startOffset match the persisted state.
> Normal consumption can advance startOffset without changing stateEpoch.
> Consequently, an identical initialization retry arriving after consumption
> has progressed is no longer recognized as a duplicate. It generates a new
> initialization snapshot, overwriting the persisted start offset and clearing
> acknowledgment batches and delivery completion count.
> h1. Trigger scenario
> When initializing newly added partitions of an already subscribed topic, the
> Group Coordinator uses startOffset=0.
> Timed-out initialization can be rescheduled while earlier persister
> initialization attempts are still retrying. If these tasks overlap within the
> same state epoch, one may complete and allow consumption to begin before
> another initialization retry arrives.
> The delayed retry can then overwrite consumption state already persisted by
> the active share partition.
> h1. Deterministic reproduction
> A regression test in ShareCoordinatorShardTest replays every generated
> coordinator record and performs the following sequence:
> 1. Initialize a share partition with stateEpoch=5 and startOffset=0.
> 2. Read its state with leaderEpoch=1.
> 3. Persist consumption progress with the same stateEpoch, startOffset=20, and
> acknowledged offsets 25–29.
> 4. Retry the original initialization request: stateEpoch=5, startOffset=0.
> 5. Read the state again.
> Expected:
> The retry preserves startOffset=20 and the acknowledgment batches.
> Actual:
> startOffset becomes 0 and the acknowledgment batches become empty.
> The attached reproduction-only patch adds:
> testInitializeStateRetryPreservesConsumptionProgress
> Run after applying the patch:
> ./gradlew :share-coordinator:test --tests
> org.apache.kafka.coordinator.share.ShareCoordinatorShardTest.testInitializeStateRetryPreservesConsumptionProgress
> -PmaxTestRetries=0 --console=plain
> h1. Impact
> The test confirms regression of the persisted share-partition start offset
> and loss of persisted acknowledgment state.
> Acknowledged records may be delivered again after the share-partition cache
> is reconstructed, for example following a leader change or restart, if the
> records are still retained and the lost state has not been restored.
> Immediate duplicate delivery is not required to trigger this issue.
> h1. Possible fix
> Treat initialization as a no-op when state already exists at the requested
> stateEpoch, regardless of startOffset. A higher stateEpoch would continue to
> allow an explicit reset.
> Current trunk increments the group epoch before an administrative share-group
> offset reset.
> Compatibility consideration: older Group Coordinator implementations,
> including 4.1.0, can perform legitimate resets without incrementing the
> epoch. The proposed change would ignore such same-epoch resets during
> mixed-version operation.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)