[
https://issues.apache.org/jira/browse/SPARK-20547?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15993426#comment-15993426
]
Shixiong Zhu commented on SPARK-20547:
--------------------------------------
Did some investigation using the reproducer. Looks like it’s not a class loader
issue. It’s because the class initialization result will be cached. My current
proposal is recreating a new class loader if a task fails.
> ExecutorClassLoader's findClass may not work correctly when a task is
> cancelled.
> --------------------------------------------------------------------------------
>
> Key: SPARK-20547
> URL: https://issues.apache.org/jira/browse/SPARK-20547
> Project: Spark
> Issue Type: Bug
> Components: Spark Shell
> Affects Versions: 2.1.0, 2.2.0
> Reporter: Shixiong Zhu
> Priority: Blocker
>
> ExecutorClassLoader's findClass may throw some transient exception. For
> example, when a task is cancelled, if ExecutorClassLoader is running, you may
> see InterruptedException or IOException, even if this class can be loaded.
> Then the result of findClass will be cached by JVM, and later when the same
> class is being loaded (note: in this case, this class may be still loadable),
> it will just throw NoClassDefFoundError.
> We should make ExecutorClassLoader retry on transient exceptions.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
