[
https://issues.apache.org/jira/browse/FLINK-16005?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17043708#comment-17043708
]
Zhenqiu Huang commented on FLINK-16005:
---------------------------------------
[~trohrmann]
I think we can leverage the HadoopConfigLoader in YarnClusterDescriptor and
also the YarnResourceManager to do the value overwrite for predefined keys
before preparing the ContainerLaunchContext.
As very limited yarn configs can be overwritten from client side. We can
predefine them in org.apache.flink.yarn.Utils. When load the Yarn Configs, we
can override the value from flink configs.
Initially, we can consider these keys:
dfs.file.replication
yarn.app.mapreduce.am.staging-dir
yarn.application.classpath
As these keys don't have the same prefix, we probably use hadoop. as prefix.
[~trohrmann][~fly_in_gis] How do you think?
How do you think?
> Propagate yarn.application.classpath from client to TaskManager Classpath
> -------------------------------------------------------------------------
>
> Key: FLINK-16005
> URL: https://issues.apache.org/jira/browse/FLINK-16005
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / YARN
> Reporter: Zhenqiu Huang
> Priority: Major
>
> When Flink users want to override the hadoop yarn container classpath, they
> should just specify the yarn.application.classpath in yarn-site.xml from cli
> side. But currently, the classpath setting can only be used in flink
> application master, the classpath of TM is still determined by the setting in
> yarn host.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
