[
https://issues.apache.org/jira/browse/YARN-2800?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14195407#comment-14195407
]
Tsuyoshi OZAWA commented on YARN-2800:
--------------------------------------
[~leftnoteasy], If we assume the labels as a configuration which can be highly
updated, ZK is not good option as you mentioned. In this case, I think
NodeLabelsManager, whose backend can be leveldb or rockdb, should be loosely
coupling with RM like TimelineServer for stabilization of RM. One option is
making NodeLabelsManager NodeLabelsServer. It means RM should work correctly
even if NodeLabelsManager is temporary unavailable. And update operation should
only affect NodeLabelsManager(it doesn't affect RM). For example, RM pulls the
label information from NodeLabelsServer periodically. RM treats the lable
information as a hint and does schedule based on label information. Even
without the information, RM should schedule apps. I think this weak consistency
approach is suitable for large-scale updating.
> Should print WARN log in both RM/RMAdminCLI side when
> MemoryRMNodeLabelsManager is enabled
> ------------------------------------------------------------------------------------------
>
> Key: YARN-2800
> URL: https://issues.apache.org/jira/browse/YARN-2800
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: client, resourcemanager
> Reporter: Wangda Tan
> Assignee: Wangda Tan
> Attachments: YARN-2800-20141102-1.patch, YARN-2800-20141102-2.patch
>
>
> Even though we have documented this, but it will be better to explicitly
> print a message in both RM/RMAdminCLI side to explicitly say that the node
> label being added will be lost across RM restart.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
