dwsmith1983 commented on PR #5613: URL: https://github.com/apache/datafusion-comet/pull/5613#issuecomment-5861613238
> While a fair-pool acquire is parked in Spark's `ExecutionMemoryPool.acquireMemory`, the executor's `comet-memory-usage-log` thread blocks in `Native.getMemoryUsage` on the same lock, because reading the pool's reservation takes it. That is the lock this PR stops holding across the Spark call. On main, `try_grow` holds `state` across `try_acquire`, and `reserved()` needs the same lock, so the logger waits for as long as the acquire is parked. Here the limit check and the charge finish under the lock, and the anchor, acquire and release calls all run after it is dropped, so `reserved()` returns straight away with the in-flight bytes counted. The greedy pool keeps its usage in an atomic and `TrackConsumersPool` only forwards the call, so no other pool type has this wait. `2987a934f` adds `reserved_is_not_blocked_by_an_acquire_parked_in_spark`, which parks an acquire in the test stub and expects `reserved()` to return within a second, including the parked task's bytes. With the lock held across the Spark call it times out. Could you approve a CI run on that 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
