hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Wangda Tan (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (YARN-4948) Support node labels store in zookeeper
Date Fri, 02 Sep 2016 00:28:22 GMT

    [ https://issues.apache.org/jira/browse/YARN-4948?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15457059#comment-15457059
] 

Wangda Tan commented on YARN-4948:
----------------------------------

Sorry for the late response, I just discussed with [~subru], and talked to [~naganarasimha_gr@apache.org]
few months before.

Thanks for taking up this JIRA, [~wjlei], I would suggest:
- If you want it works with WASB file system (or other file system doesn't support append),
you can try to use NonAppendableFSNodeLabelStore.  
- If you don't want YARN fails because of HDFS failures, you can setup yarn.node-labels.fs-store.retry-policy-spec

- Reasonable number of nodes and reasonable frequency of node label updating is already supported
by existing stores. There will be a lot of issue if #nodemanager goes beyond 20000+, for example,
scheduler itself becomes a bottleneck.

I would say adding the new ZK store impl will add extra overhead to code maintenance. And
since pluggable node label store is supported, if you really want to run a ZK store, you can
configure it in yarn-site.xml. So I suggest to keep this open until this becomes a common
requirement for node label.

> Support node labels store in zookeeper
> --------------------------------------
>
>                 Key: YARN-4948
>                 URL: https://issues.apache.org/jira/browse/YARN-4948
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: resourcemanager
>            Reporter: jialei weng
>            Assignee: jialei weng
>         Attachments: YARN-4948.001.patch, YARN-4948.002.patch, YARN-4948.003.patch, YARN-4948.006.patch,
YARN-4948.007.patch
>
>
> Support node labels store in zookeeper. The main scenario for this is to give a way to
decouple yarn with HDFS. Since nodelabel is a very important data for yarn, if hdfs down,
yarn will fail to start up,too. So it is meaningful for make yarn much independence when user
serve both yarn and HDFS. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org


Mime
View raw message