bbejeck opened a new pull request, #22478: URL: https://github.com/apache/kafka/pull/22478
In 4.3, KIP-1035 moved the changelog offset into RocksDB and removed the forced flush on commit, so the persisted offset is now only made durable by an organic memtable flush or a clean close. When that offset goes stale — after an unclean exit, or a clean shutdown followed by changelog truncation/compaction while the instance is down — and the changelog log-start offset has advanced past it, the restore consumer seeks out of range and throws `OffsetOutOfRangeException`, which Streams converts to a TaskCorruptedException (full local-state wipe and rebuild). This happens far more often than in 4.2 (where the forced flush kept the offset within roughly commit.interval.ms), affecting both at-least-once and exactly-once and hitting windowed/segmented stores hardest. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
