davidzollo commented on PR #11757: URL: https://github.com/apache/seatunnel/pull/11757#issuecomment-5549209322
## Closed the FAILED-fallthrough gap Daniel flagged Pushed `4336d46`, which fixes the gap in Issue 1: `finishOwnedResources()` (the `completionLatch==0` path) already checks that a tracker still owns the active `TaskGroupContext` before touching `taskAsyncFunctionFuture`/`timerFlushFutures`, but `cancelAllTask()` had no such guard — and it's reachable two other ways that skip that check entirely: 1. `taskDone()`'s FAILED-fallthrough, whenever *any* task in the group fails while the group isn't cancelled — including when `completionLatch` hasn't reached zero yet, i.e. a sibling task can still fail after this tracker's own generation has already lost the ownership race. 2. The `cancellationFuture.whenComplete()` handler installed in the constructor. In both cases a stale tracker would unconditionally cancel and remove whatever `taskAsyncFunctionFuture`/`timerFlushFutures` entries currently sit under its `TaskGroupLocation` — which, after a restore, belong to the newer generation. Same corruption class this PR exists to fix, just a different call path than the one already covered. **Fix:** `cancelAllTask()` now takes the same `synchronized (TaskExecutionService.this)` ownership check `finishOwnedResources()` already uses, guarding only the two shared, location-keyed calls (`cancelAsyncFunction`, `cancelTimerFlushForTaskGroup`). The per-tracker-private `blockingFutures`/`currRunningTaskFuture` cancellation above it is untouched, since those are never shared across generations. Fixing it centrally in `cancelAllTask()` covers both call sites in one change rather than patching `taskDone()` alone. **Test:** added `testStaleFailedTaskDoneDoesNotCleanupNewerGenerationResources`, using a two-task old-generation group and failing only the first task so `completionLatch` never reaches zero — isolating the FAILED-fallthrough path from `finishOwnedResources()`'s already-guarded branch, per Daniel's exact critique of the existing test (`testStaleTaskDoneDoesNotCleanupNewerGenerationResources` uses a single always-successful task and never exercises this path). @DanielLeens — over to you for a fresh from-scratch review of the cleaned head. -- 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]
