You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Ray Chiang (JIRA)" <ji...@apache.org> on 2016/02/23 20:26:18 UTC

[jira] [Commented] (YARN-3607) Allow users to choose between failing the daemons vs failing the apps/containers

    [ https://issues.apache.org/jira/browse/YARN-3607?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15159488#comment-15159488 ] 

Ray Chiang commented on YARN-3607:
----------------------------------

Two suggestions:

1) Since this is a setting that affects all daemons, it makes sense to have one setting per daemon type, such as yarn.resourcemanager.fail-fast and yarn.nodemanager.fail-fast.

2) There is going to be a lot of places in the YARN code where this variable could be checked.  I'm thinking the first task/subtask would be to just add the variable definitions now and then let the functionality be added where it's appropriate.


> Allow users to choose between failing the daemons vs failing the apps/containers
> --------------------------------------------------------------------------------
>
>                 Key: YARN-3607
>                 URL: https://issues.apache.org/jira/browse/YARN-3607
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: nodemanager, resourcemanager, scheduler
>    Affects Versions: 2.7.0
>            Reporter: Karthik Kambatla
>            Assignee: Ray Chiang
>
> We often run into cases where we are faced with the option of failing the daemon (fail-fast) vs failing user's work and keep the cluster running. There is no clear right way to handle these situations - some users would like to be conservative and let the daemons run, while others would like to fail-fast. 
> Today, we handle these case-by-case and go by what the people working on it feel is the right way to handle things. Examples include how we handle app recovery failures, queue-changes on RM restart. 
> Users should be able to choose between these two extremes, and have all these situations handled the same way. 



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