[
https://issues.apache.org/jira/browse/SAMZA-2619?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17318213#comment-17318213
]
Lakshmi Manasa Gaduputi commented on SAMZA-2619:
------------------------------------------------
*Problem:* the patch [PR#1480|https://github.com/apache/samza/pull/1480] does
not work: throws NPE as appCtx.getAMContainerResourceRequest() returns null and
is null till the end of the ClientHelper.submitApplication method. As in,
AMContainerResourceRequest will always remain null within
ApplicationSubmissionContext while in ClientHelper.
*NPE Reason:* Peeking into yarn code shows that AMContainerResourceRequest is
created, if null within
[RMAppManager.validateAndCreateResourceRequest|https://github.com/apache/hadoop/blob/release-2.7.1/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-resourcemanager/src/main/java/org/apache/hadoop/yarn/server/resourcemanager/RMAppManager.java#L372]
by using the mem,cpu set in the ApplicationSubmissionContext.setResource.
*YARN code into samza?*: The only way to set node label only for AM during
submit application is to use
[setAMContainerResourceRequest.setNodeLabelExpression in
ApplicationSubmissionContext|https://hadoop.apache.org/docs/r2.7.3/hadoop-yarn/hadoop-yarn-site/NodeLabel.html].
However, creating AMContianerResourceRequest within samza's ClientHelper will
require AM_CONTAINER_PRIORITY constant which is a part of yarn-server package
NOT api, common or client which samza uses. Pulling yarn-server package into
samza is seems plain WRONG! Additionally, it also means in some sense
replicating what RMAppManager.validateAndCreateResourceRequest does to create
the AMContainerRequest.
*Workaround (tested and works)* - is in the samza-job, use
yarn.am.container.label=desired-label AND yarn.container.label="" -- relying on
this javadoc in yarn for[
SubmissionApplicationContext|https://github.com/apache/hadoop/blob/release-2.7.1/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-api/src/main/java/org/apache/hadoop/yarn/api/records/ApplicationSubmissionContext.java#L437]
and this comment from YARN-2493. This workaround places AM on desired-label
node while containers go to general nodes without a label.
> yarn.am.container.label is applied to entire samza job (all its containers)
> instead of only the AM.
> ---------------------------------------------------------------------------------------------------
>
> Key: SAMZA-2619
> URL: https://issues.apache.org/jira/browse/SAMZA-2619
> Project: Samza
> Issue Type: Bug
> Reporter: Lakshmi Manasa Gaduputi
> Assignee: Lakshmi Manasa Gaduputi
> Priority: Major
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> yarn.am.container.label is applied to entire samza job (all its containers)
> instead of only the AM.
> Yarn supports 3 forms of applying node label to containers – am only,
> container only and entire job.
> [https://hadoop.apache.org/docs/current/hadoop-yarn/hadoop-yarn-site/NodeLabel.html#Specifying_node_label_for_application
> ]Samza's ClientHelper.scala finds the am container label and applies it to
> the application submission context which applies the label to the entire job
> instead of only the am.
> this restricts jobs which prefer to have label only for am or different
> labels for am and containers.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
