DanielLeens commented on issue #12456: URL: https://github.com/apache/seatunnel/issues/12456#issuecomment-6059630871
Classification: A / Zeta resource-management and class-loader lifecycle. This remains a P0 stability investigation because repeated job cleanup has a demonstrated Metaspace-OOM path. @wangbing505 thank you for retesting #12626 and supplying the new heap paths. These are not the two roots that PR currently cleans: its cleaner targets the Hadoop `WritableComparator` registry and the MongoDB buffer-pool pruner. It does not establish cleanup ownership for the Hadoop IPC Parameter Sending Thread, ServiceLoader token renewers, `CodecPool`, or the JNA/finalizer chain. #12626 must therefore be treated as partial, not as a general class-loader-leak fix. Please add, for the exact #12626 commit and the smallest reproducer, a post-job thread dump showing whether the IPC thread is alive, its context class loader, and the owning Hadoop client/RPC object; plus the shortest complete retaining path to a GC root for each remaining loader. Include the connector JAR set, Hadoop/MongoDB versions, number of submit/cancel cycles, and matched before/after heap dumps. That will tell us whether an owner can be safely closed. Do not add blanket reflection over JVM finalizers, global Hadoop caches, or arbitrary threads. Any follow-up must be one root owner at a time with a real-library regression, preserve still-running jobs and classloader-cache mode, and show that the released loader becomes collectable after the job ends. No label or assignment changed here. -- 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]
