The GitHub Actions job "Flink Connector Kafka CI" on flink-connector-kafka.git/fix-committer-abort-fenced-reused-id has failed. Run started by GitHub user slachiewicz (triggered by slachiewicz).
Head commit for run: 7fce0b8b6926ef8e56bc10933336c9eb42abf3e1 / Sylwester Lachiewicz <[email protected]> [FLINK-40626][Connectors/Kafka] fix: Abort superseded precommitted transactions on recovery With POOLING, a committed transactional id is reused for a later checkpoint under a bumped epoch. A recovery from the earlier checkpoint lists the id as precommitted, so LISTING skipped it, the committer's commit was fenced, and the newer transaction stayed open until transaction.timeout.ms, pinning the last stable offset. The writer state now records producer id and epoch per precommitted transaction (serializer v3; v2 reads as unknown and keeps the skip; earlier connector versions cannot read v3, so a rollback needs a savepoint taken before the upgrade). On recovery, LISTING describes the precommitted ids and aborts one the broker still reports ONGOING under a newer epoch or a different producer id: that transaction belongs to the failed attempt or a zombie, and the newest attempt is entitled to fence it. The committer is unchanged, so a fenced commit from an older attempt can never fence a live owner. The aborted id stays tracked as ongoing. The committer's fenced commit of the restored transaction reports the id as finished through the backchannel, and the writer drains that only at its next snapshot; had POOLING already reused the id for the first new transaction, the stale report would close the new producer after prepareCommit handed it to the committer and lose that checkpoint (KafkaSinkRecoveryITCase). The id is released by that report, or by the sweep once a later checkpoint's transaction finishes. Generated-by: Claude Code (Claude Opus 5, Claude Sonnet 5, Claude Opus 5.5) Report URL: https://github.com/apache/flink-connector-kafka/actions/runs/39342878234 With regards, GitHub Actions via GitBox
