[
https://issues.apache.org/jira/browse/FLINK-12620?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17236826#comment-17236826
]
Mike Kaplinskiy commented on FLINK-12620:
-----------------------------------------
We discovered a workaround to this problem: passing
{{-XX:+AlwaysLockClassLoader}} to the job manager JVMs. This way the fix
doesn't require a custom compiled version of Flink with the patch above.
> Deadlock in task deserialization
> --------------------------------
>
> Key: FLINK-12620
> URL: https://issues.apache.org/jira/browse/FLINK-12620
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Task
> Affects Versions: 1.8.0
> Reporter: Mike Kaplinskiy
> Priority: Major
> Attachments: BatchJob.java, jstack_repro.txt, jstack_snippet.txt
>
>
> When running a batch job, I ran into an issue where task deserialization
> caused a deadlock. Specifically, if you have a static initialization
> dependency graph that looks like this (these are all classes):
> {code:java}
> Task1 depends on A
> A depends on B
> B depends on C
> C depends on B [cycle]
> Task2 depends on C{code}
> What seems to happen is a deadlock. Specifically, threads are started on the
> task managers that simultaneously call BatchTask.instantiateUserCode on both
> Task1 and Task2. This starts deserializing the classes and initializing them.
> Here's the deadlock scenario, as a stack:
> {code:java}
> Time---->
> T1: [deserialize] -> Task1<clinit> -> A<clinit> -> B<clinit> -> (wait for
> C<clinit>)
> T2: [deserialize] -> Task2<clinit> -> C<clinit> -> (wait for
> B<clinit>){code}
>
> A similar scenario from the web:
> [https://www.farside.org.uk/201510/deadlocks_in_java_class_initialisation] .
>
> For my specific problem, I'm running into this within Clojure -
> {{clojure.lang.RT}} has a dep on {{clojure.lang.Util}} which has a dep with
> {{clojure.lang.Numbers}} which depends on {{clojure.lang.RT}} again.
> Deserializing different clojure functions calls one or the other first which
> deadlocks task managers.
>
> I built a version of flink-core that had
> {{org.apache.flink.util.InstantiationUtil.readObjectFromConfig}}
> synchronized, but I'm not sure that it's the proper fix. I'm happy to submit
> that as a patch, but I'm not familiar enough with the codebase to say that
> it's the correct solution - ideally all Java class loading is synchronized,
> but I'm not sure how to do that.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
