hwanju commented on issue #11196: [FLINK-16246][connector kinesis] Exclude AWS SDK MBean registry from Kinesis build URL: https://github.com/apache/flink/pull/11196#issuecomment-591253938 Although I am also not sure what would be the ramification of excluding`SdkMBeanRegistrySupport` to other depending classes, it seems like no-op replacement seems safe and clever. Related to `FlinkKinesisProducer#close()` solution, would it be viable to include `AwsSdkMetrics#unregisterMetricAdminMbean()` inside `BlobLibraryCacheManager#releaseClassLoader` to get that unregistered once all the tasks on the job are closed? But it seems too specific call. At any rate, we may not make sure that removing this can resolve the leak, right? (if I am reading FLINK-16142 correctly). AFAIR, KPL also has statically launched daemon thread (file age manager), which is not shut down on task close and holds on to the class loaded by user class loader, leading to leak. In this case, not sure how just removing this MBean registry would get things better. We've seen OOM killed with unlimited metaspace, but it seems like it hits metaspace OOM before eating up all the memory as of 1.10 (which we have not run yet). Curious if you have measured number of class loaders before/after the fix.
---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
