[
https://issues.apache.org/jira/browse/FLINK-10904?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16693492#comment-16693492
]
Stephan Ewen commented on FLINK-10904:
--------------------------------------
Flink currently instantiates one classloader per parallel task that gets
created, so across different tasks you have different classloaders, even if
running on the same TaskManager.
> Expose Classloader before Pipeline execution
> --------------------------------------------
>
> Key: FLINK-10904
> URL: https://issues.apache.org/jira/browse/FLINK-10904
> Project: Flink
> Issue Type: Wish
> Reporter: Luka Jurukovski
> Priority: Minor
>
> Not sure if this is something that I just have to deal with. However it would
> be nice if the classloader can be accessed before the pipeline starts
> executing. The case for this is that I am loading classes that contain Flink
> operators. I am running into classdef not found issues because the
> classloader used by Flink is different then the application program that is
> being run. Currently what I have been doing as a work around is adding the
> libs that cause these issues in the /lib directory of the Flink cluster and
> marking it as provided in the application jar that is uploaded to the Flink
> cluster. The issues with this are two fold, first it makes deployment more
> complex, as well as there are cases where Classloading causes exceptions to
> arise in unusual circumstances. Ie RabbitMQ connector caused issues only when
> it was auto-recovering the connection, but not during normal ingest. I can
> elaborate further if needed
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
