MartijnVisser opened a new pull request, #29553:
URL: https://github.com/apache/flink/pull/29553

   Unchanged backport of https://github.com/apache/flink/pull/27157 
(9b21a9716cc), which went to master, release-1.19, release-2.0 and release-2.1 
but not to this branch. It is the release-1.19 commit bd916bd89e7 (RocksDB 
only, ForSt on 1.20 has no snapshot strategy) plus the hotfix 1e367509833 that 
moves the "Will re-use" log line to TRACE.
   
   When the completion notification of checkpoint N arrives after N+1 has 
started, N+1 re-uploads the sst files of N. Once N+1 completes the JobManager 
drops the handles of N, N+2 still references them, and the 
`SharedStateRegistry` rejects that acknowledgement with "Attempt to reference 
unknown state". The acknowledgement is lost and the checkpoint waits for its 
timeout. The fix removes every re-uploaded file from the base set.
   
   On release-1.20 this is what fails `WindowDistinctAggregateStressITCase` 
(FLINK-40663), 37 of 243 table jobs on the GitHub Actions nightlies and pushes 
since 2026-09-14. Last seen in 
https://github.com/apache/flink/actions/runs/39531020653/job/119010155405 
(nightly 2026-10-21, `Java 17 / Test (module: table)`). FLINK-35499 recorded 
the same exception on 1.20 in July 2024.
   
   `testCheckpointIsIncrementalWithLateNotification` fails on this branch 
without the production change and passes with it, on JDK 8 and 17. The ITCase 
under CI's flags in a two-CPU container hit the exception in 1 of 30 runs on 
the base and in 0 of 89 with this change. The fix's prune trace is in 83 of 
those 89 logs. The rocksdb module tests, `EventTimeWindowCheckpointingITCase`, 
spotless, checkstyle and the planner ArchUnit tests pass.
   
   ---
   
   ##### Was generative AI tooling used to co-author this PR?
   
   - [X] Yes (please specify the tool below)
   
   Generated-by: Claude Code (Claude Fable 5.1)
   


-- 
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]

Reply via email to