[
https://issues.apache.org/jira/browse/FLINK-12152?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16821050#comment-16821050
]
vinoyang commented on FLINK-12152:
----------------------------------
[~Paul Lin] Sorry for the late reply, Our scene just like I described in the
Description. Our job's parallelism is very large, when the job is started, we
can not open the Flink web UI, but we can see the job manager's log on YARN
console is still printing and see many Akka timeout log and the job be
restarted. So we infer that the Akka's inner thread pool is busy and the AM(JM)
need more vcores. Sorry, there is no log message, it happened a few days ago.
Moreover, this is also a configurable option in Spak. cc [~till.rohrmann]
> Make the vcore that Application Master used configurable for Flink on YARN
> --------------------------------------------------------------------------
>
> Key: FLINK-12152
> URL: https://issues.apache.org/jira/browse/FLINK-12152
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / YARN
> Reporter: vinoyang
> Assignee: vinoyang
> Priority: Major
>
> Now, for Flink on YARN deployment mode, each am's vcores is specified to 1
> (hard code).
> In some scene, we found many Akka timeout logs, the Flink web UI cannot be
> opened, but it is alive. I think there is no more threads resource to be used
> for am. So we suggest that make the vcores num of application master can be
> configurable.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
