[ 
https://issues.apache.org/jira/browse/YARN-7863?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16534148#comment-16534148
 ] 

Sunil Govindan commented on YARN-7863:
--------------------------------------

Thanks [~Naganarasimha] for the comments. 
{quote}bq. 1.who is doing the validation of the resource request?
{quote}
>> Validation will be done by scheduler module similar to validate 
>> ResourceRequest. 
{quote}bq. 2. If there are any open resource request for specified attribute, 
do we allow removing of that attribute from Attribute Manager ?
{quote}
>> I think this is not required for now. We can see if any attributes are used, 
>> and stop deleting such as its already in used. For pending case, some user 
>> has asked for some resources, but scheduling didnt happen as NMs were down 
>> or all nodes are full with that attribute. For latter case, we wont delete 
>> the attribute as its used. In case first case, its possible that NMs were 
>> down and may come back late or we dont know the timeline. At that time, it 
>> makes sense to remove. However we can think more corner cases and come with 
>> some protection later.
{quote}bq. 3 .Confused little with the logic (may be earlier code to yours), i 
was always of the opinion that the partition is the first to be compared and 
then with tag or attribute here
{quote}
>> Actually attribute and tags are different. We can have a combined OR or AND 
>> across attributes/tags however these two need not have to be considered in 
>> same expression as use case is specifically for attribute or tag. I 
>> discussed with [~leftnoteasy] on this, and he also suggested to keep 
>> attribute different to tag and make the relation by using OR and AND.
{quote}bq. PlacementConstraintsUtil, ln no 145-156 : Scheduler node needs to be 
having map of NodeAttributeKey -> NodeAttribute value to make it O(1), as well 
this will more useful when we go other types and operator based comparison 
happens. Or to consider for current scenario atleast it would be better to 
override NodeAttribute's equals method.
{quote}
>> Sure. I will update similar.

I am improving this patch with some test cases and will share a new one 
tomorrow.

> Modify placement constraints to support node attributes
> -------------------------------------------------------
>
>                 Key: YARN-7863
>                 URL: https://issues.apache.org/jira/browse/YARN-7863
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>            Reporter: Sunil Govindan
>            Assignee: Sunil Govindan
>            Priority: Major
>         Attachments: YARN-7863.v0.patch
>
>
> This Jira will track to *Modify existing placement constraints to support 
> node attributes.*



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

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

Reply via email to