DanielLeens commented on issue #13137:
URL: https://github.com/apache/seatunnel/issues/13137#issuecomment-6687112162

   Thanks @SEPURI-SAI-KRISHNA — this is the missing reproducible evidence. It 
changes the status from an unconfirmed aggregation-scope hypothesis to a 
reproducible realtime-observability defect.
   
   I checked current `dev` (`9b8252eaf0c0eed5d80d5fd10a3b15eafeab9f58`). 
`SeaTunnelTask.init()` reloads the metrics snapshot for each `TaskLocation`, 
while `TaskGroupWithIntermediateBlockingQueue.getQueueCache()` creates the 
`IntermediateQueueSize`, per-queue size, blocked-time, and capacity counters 
only in the `MetricsContext` supplied by whichever task reaches the shared 
queue first. That ownership can therefore change across restored attempts, 
leaving a previous task-local queue row in the snapshot while another task 
registers the live queue. This is consistent with the observed extra 
zero-valued rows and the lowered aggregate ratios.
   
   I also checked the current open PR set: #13133 and #13136 address job-info 
metadata after restore, not these task-local intermediate-queue counters; I 
found no open PR for this exact defect. @JayModha100, the original reporter did 
not claim a PR and @SEPURI-SAI-KRISHNA has said they will not open one, so the 
focused reproduction/fix scope is available.
   
   For a PR, please keep the boundary narrow:
   
   1. Reproduce a deterministic owner change across same-ID restore attempts 
and assert the live realtime-edges response counts only the current 
queue-bearing groups, with the matching total capacity and ratio denominator.
   2. Clear or replace only stale intermediate-queue metric keys when a 
task-local context is restored; do not clear unrelated task metrics or broadly 
discard the metrics snapshot.
   3. Treat `--get_running_job_metrics` history as a separate API contract. The 
proposed live-counter fix must not silently redefine that historical view.
   4. Include the sanitized FakeSource configuration and before/after raw edge 
responses (or equivalent deterministic test fixtures) in the PR description, 
and cover both blocking and disruptor queue paths if they share the 
registration mechanism.
   
   Please link the PR here once the regression test proves the invariant. No 
labels or assignments were changed.


-- 
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