[
https://issues.apache.org/jira/browse/YARN-1639?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13889978#comment-13889978
]
Xuan Gong commented on YARN-1639:
---------------------------------
bq. After this patch, if there are 4 RM's in HA setup, then do we have to
enumerate all RM instance configs in the conf file. ie. for RM Admin Service
RPC port we need to have 4 entries suffixed by each RM id (the values may be
identical)? Was this the case earlier too? Or earlier, we could have specified
an single config for RM Admin Service RPC and each RM would use its rm id to
create a suffixed entry in the config from the single config.
Before this patch, if we want to config HA, we need to have
yarn.resourcemanager.ha.rm-ids : rm1, rm2, rm3, rm4
And we need to set yarn.resourcemanager.ha.id to rm1 or rm2 or rm3 or rm4 for
four different configurations
Then we should provide all RM rpc addresses suffixed by each RM id for all four
different configurations.
After this patch,
we do not need to provide any yarn.resourcemanager.ha.id, other configurations
are the same. In that case, we can share the same configuration for the
different machines.
So, we do not provide yarn.resourcemanager.ha.id, and the rm figures out its
own rm_id by comparing its local address with RM_Address. These is one thing we
need to pay attention is that we can not launch more than one rm in one single
machine. If we really want to do that (for example, do the testing,
miniYarnCluster), we need to set values for yarn.resourcemanager.ha.id
> YARM RM HA requires different configs on different RM hosts
> -----------------------------------------------------------
>
> Key: YARN-1639
> URL: https://issues.apache.org/jira/browse/YARN-1639
> Project: Hadoop YARN
> Issue Type: Sub-task
> Components: resourcemanager
> Reporter: Arpit Gupta
> Assignee: Xuan Gong
> Fix For: 2.4.0
>
> Attachments: YARN-1639.1.patch, YARN-1639.2.patch, YARN-1639.3.patch,
> YARN-1639.4.patch, YARN-1639.5.patch
>
>
> We need to set yarn.resourcemanager.ha.id to rm1 or rm2 based on which rm you
> want to first or second.
> This means we have different configs on different RM nodes. This is unlike
> HDFS HA where the same configs are pushed to both NN's and it would be better
> to have the same setup for RM as this would make installation and managing
> easier.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
