[ 
https://issues.apache.org/jira/browse/YARN-5478?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15941628#comment-15941628
 ] 

Naganarasimha G R commented on YARN-5478:
-----------------------------------------

Thanks [~wangda] for providing the patch and the description of the approach at 
the high level for 6345, 1042 and 3409.
By ??Node attributes: YARN-4902?? i presume you were referring to YARN-3409. 
Overall idea of having new resource request seems to fit in very good but had 
few concerns/queries with the structure of PlacementStratergy.
1. IIUC NodeAttributes gets filled in *Affinity target* of Placement strategy. 
And as i understand based on example we directly take in the label expression 
which can contain *\!<label>* or *<label>\!=<value>*. So it kind of contradicts 
the name *Affinity*.
2. *PlacementStrategy* has scope which takes in the value of *NODE/RACK*. But 
*NodeAttributes* makes sense only for a Node. And it would have not much 
significance for a rack. For ex, i would not like to place a container in rack 
which has one node with GPU but would like to place in that node itself.

So i feel that struture fits in fine for allocation tag but not NodeAttributes. 
Will think more on how to place it better in PlacementStrategy.

> [YARN-4902] Define Java API for generalized & unified scheduling-strategies.
> ----------------------------------------------------------------------------
>
>                 Key: YARN-5478
>                 URL: https://issues.apache.org/jira/browse/YARN-5478
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Wangda Tan
>            Assignee: Wangda Tan
>         Attachments: YARN-5478.1.patch, YARN-5478.2.patch, 
> YARN-5478.preliminary-poc.1.patch, YARN-5478.preliminary-poc.2.patch
>
>
> Define Java API for application to specify generic scheduling requirements 
> described in YARN-4902 design doc.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to