DanielLeens opened a new issue, #12947:
URL: https://github.com/apache/seatunnel/issues/12947

   ## Bug description
   
   During Zeta fixed-slot scheduling and failover convergence, a multi-slot 
request can partially succeed before another slot request fails. The current 
request path completes the allocation failure before asynchronous rollback 
finishes. A later retry can then race with slots that are still owned by the 
failed request.
   
   The same failure chain exposes several related state inconsistencies:
   
   - missing-slot retries compress request indexes and can overwrite previously 
allocated non-contiguous slots;
   - a resource that fails fixed-worker pre-check is not reliably carried into 
dynamic-worker fallback;
   - synchronous slot-release failures can escape the rollback future and leave 
the outer request unresolved;
   - failed task deployment can transiently publish RUNNING, and the pipeline 
can publish RUNNING without checking every vertex result;
   - null, exceptional, or cancelled slot futures can bypass complete cleanup 
bookkeeping;
   - querying job status during the live-to-history convergence window can 
dereference a missing historical state.
   
   ## Expected behavior
   
   - Preserve each original request index through fixed and dynamic allocation 
attempts.
   - Complete a failed allocation only after bounded rollback attempts finish, 
and retain rollback failure as diagnostic context.
   - Retry only releases that have not already succeeded.
   - Publish task and pipeline RUNNING states only after successful deployment, 
while accepting vertices that have already finished normally.
   - Treat null, exceptional, and cancelled slot futures as missing allocations 
while retaining successfully owned slots for cleanup.
   - Return UNKNOWABLE when neither live nor historical job status is available.
   
   ## Scope
   
   This issue does not include the stale-slot pipeline restore problem tracked 
by #12509 and addressed by #12510. It also does not include the CDC 
acknowledgement chain in #12163 or the idle-reader checkpoint recovery in 
#12388.


-- 
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