[jira] [Commented] (FLINK-20614) Registered sql drivers not deregistered after task finished in session cluster

["Chesnay Schepler (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-20614?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17250028#comment-17250028 ] 

Chesnay Schepler commented on FLINK-20614:
------------------------------------------

This issue was documented in FLINK-19005 (https://ci.apache.org/projects/flink/flink-docs-release-1.12/ops/debugging/debugging_classloading.html#unloading-of-dynamically-loaded-classes-in-user-code).

The tomcat approach pretty much does what I suggested in FLINK-19005. Looking at what they [actually have to do to make it work |https://github.com/apache/tomcat/blob/efc6af6778ff3c1605d8b053f6fd2a4d9fd8e0d3/java/org/apache/catalina/loader/WebappClassLoaderBase.java#L1673] I'd rather not actually go down that route.

> Registered sql drivers not deregistered after task finished in session cluster
> ------------------------------------------------------------------------------
>
>                 Key: FLINK-20614
>                 URL: https://issues.apache.org/jira/browse/FLINK-20614
>             Project: Flink
>          Issue Type: Bug
>          Components: Connectors / JDBC, Runtime / Task
>    Affects Versions: 1.12.0, 1.13.0
>            Reporter: Kezhu Wang
>            Priority: Major
>
> {{DriverManager}} keeps registered drivers in its internal data structures which prevents they from gc after task finished. I confirm it in standalone session cluster by observing that {{ChildFirstClassLoader}} could not be reclaimed after several {{GC.run}}, it should exist in all session clusters.
> Tomcat documents [this|https://ci.apache.org/projects/tomcat/tomcat85/docs/jndi-datasource-examples-howto.html#DriverManager,_the_service_provider_mechanism_and_memory_leaks] and fixes/circumvents this with [JdbcLeakPrevention|https://github.com/apache/tomcat/blob/master/java/org/apache/catalina/loader/JdbcLeakPrevention.java#L30].
> Should we solve this in runtime ? Or treat it as connector and clients' responsibility to solve it using {{RuntimeContext.registerUserCodeClassLoaderReleaseHookIfAbsent}} or similar ?
> Personally, it would be nice to solve in runtime as a catch-all to avoid memory-leaking and provide consistent behavior to clients cross per-job and session mode.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)