[
https://issues.apache.org/jira/browse/FLINK-10352?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16619601#comment-16619601
]
Stephan Ewen commented on FLINK-10352:
--------------------------------------
This is sensitive core change and behavior breaking. This needs to have a
design first.
Some issues that come immediately to mind to clarify first:
- This changes behavior of static variables across tasks and across
re-executions, can hence break programs.
- Will the old model still be supported for applications that relied on that?
What will the config switch be?
- How how do we deal with (unexpected but possible) different sets of
libraries attached to tasks?
> Bind Task's URLClassLoader to TaskExecutor's slot
> -------------------------------------------------
>
> Key: FLINK-10352
> URL: https://issues.apache.org/jira/browse/FLINK-10352
> Project: Flink
> Issue Type: Improvement
> Components: Distributed Coordination
> Affects Versions: 1.5.3, 1.6.0, 1.7.0
> Reporter: Till Rohrmann
> Assignee: Jiayi Liao
> Priority: Major
> Fix For: 1.7.0
>
>
> The {{Task's}} {{URLClassLoader}} is used to dynamically load user code
> classes. The lifetime of this class loader is bound to the {{Task}} and,
> thus, it can happen that we generate multiple class loader for the same job
> if multiple {{Tasks}} are deployed to the same slot successively.
> We could avoid this situation by binding the lifetime of the
> {{URLClassLoader}} to the slot instead of the {{Task}}. The slot will only be
> released if the {{JobMaster}} no longer needs it. That way we won't
> potentially recreate the {{URLClassLoader}} for every submitted {{Task}}.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
