dbtsai commented on PR #56377: URL: https://github.com/apache/spark/pull/56377#issuecomment-4654483153
Thanks both for the careful reviews. Addressed in the latest push (`b9c27909f20`): **@viirya** — reworded the race-correctness comment. It no longer attributes safety to the symmetric store-then-load on the volatile `closedTimeMs` (which, as you note, is StoreLoad/Dekker-vulnerable in isolation). It now points at `ConcurrentHashMap`'s per-key linearizability: if the cleanup `forEach` misses our freshly `compute`d entry, then our insert linearized *after* `close()`'s read of that key's bin; since `closedTimeMs` is written before that bin read (program order in `close()`) and our `compute()` precedes the `isClosing` read (program order here), transitivity forces the recheck to observe `isClosing == true`. Also dropped the "forEach reflects entries already present when it starts" wording since `forEach` is weakly consistent. **@cloud-fan** — good catch, that's the same bug class. A foreachBatch runner cleaner registered (planner ~L3594) after `close()` has passed `cleanUpAll()` (L404) can be missed by that reaper, and if my async `stop()` terminates the query before the cleaner's `onQueryTerminated` listener is in place, that reaper misses it too — stranding the Python worker. I added the symmetric guard to `CleanerCache.registerCleanerForQuery`: it re-checks `sessionHolder.isClosing` after inserting and cleans the runner up immediately if the session is closing, plus a regression test. The query strand was the confirmed production symptom; this closes the analogous runner strand in the same window. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
