[ 
https://issues.apache.org/jira/browse/YARN-3542?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15044562#comment-15044562
 ] 

Sidharta Seethana commented on YARN-3542:
-----------------------------------------

Given that one of the goals of this JIRA/patch is to deprecate the current 
implementation, how does this approach sound? 

 * Move the implementation from {{CgroupsLCEResourcesHandler}} into the new 
resource handler. Mark the former as deprecated.
 * The existing config setting remains for backward compatibility - when 
{{CgroupsLCEResourcesHandler}} is specified as the resource handler class, use 
the new resource handler implementation ( handled in {{ResourceHandlerModule}} 
) . The existing integration in {{LinuxContainerExecutor}} can probably be 
removed at this point. 
 * Config behavior : If the existing config setting and the new config setting 
are both explicitly specified and are incompatible, bail out ( this might be a 
bit tricky depending on how defaults are handled for these settings )

A couple of other minor comments based on a first look : 
 * LOG.debug() statements should be guarded by a config check, IMO. 
 * YarnConfiguration changes : Disk/Network configs define a prefix ( e.g 
NM_DISK_RESOURCE_PREFIX ) - maybe something similar should be used for CPU ? 

> Re-factor support for CPU as a resource using the new ResourceHandler 
> mechanism
> -------------------------------------------------------------------------------
>
>                 Key: YARN-3542
>                 URL: https://issues.apache.org/jira/browse/YARN-3542
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: nodemanager
>            Reporter: Sidharta Seethana
>            Assignee: Sidharta Seethana
>            Priority: Critical
>         Attachments: YARN-3542.001.patch, YARN-3542.002.patch
>
>
> In YARN-3443 , a new ResourceHandler mechanism was added which enabled easier 
> addition of new resource types in the nodemanager (this was used for network 
> as a resource - See YARN-2140 ). We should refactor the existing CPU 
> implementation ( LinuxContainerExecutor/CgroupsLCEResourcesHandler ) using 
> the new ResourceHandler mechanism. 



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

Reply via email to