[ 
https://issues.apache.org/jira/browse/YARN-8459?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16524057#comment-16524057
 ] 

Wangda Tan commented on YARN-8459:
----------------------------------

[~cheersyang], 

According to our current locking design of CapacityScheduler:
1) Add/remove node/app requires CS lock. 
2) Allocate/release container acquires app/node/queue lock only for better 
performance. 

The simplest solution is to put allocate/release container under CS lock, but 
it will cause performance regression. Adding a stopping flag to app/node seems 
like cleanest solution in my mind, please share if you have any better idea.

[~sunilg], 

My intention is to put the setStopping under app/node lock instead of using 
volatile. We don't want a node is allocating container but the other thread is 
trying to remove the node.

> Capacity Scheduler should properly handle container allocation on app/node 
> when app/node being removed by scheduler
> -------------------------------------------------------------------------------------------------------------------
>
>                 Key: YARN-8459
>                 URL: https://issues.apache.org/jira/browse/YARN-8459
>             Project: Hadoop YARN
>          Issue Type: Bug
>          Components: capacity scheduler
>    Affects Versions: 3.1.0
>            Reporter: Wangda Tan
>            Assignee: Wangda Tan
>            Priority: Blocker
>         Attachments: YARN-8459.001.patch
>
>
> Thanks [~gopalv] for reporting this issue. 
> In async mode, capacity scheduler can allocate/reserve containers on node/app 
> when node/app is being removed ({{doneApplicationAttempt}}/{{removeNode}}).
> This will cause some issues, for example.
> a. Container for app_1 reserved on node_x.
> b. At the same time, app_1 is being removed.
> c. Reserve on node operation finished after app_1 removed 
> ({{doneApplicationAttempt}}). 
> For all the future runs, the node_x is completely blocked by the invalid 
> reservation. It keep reporting "Trying to schedule for a finished app, please 
> double check" for the node_x.
> We need a fix to make sure this won't happen.



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

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org

Reply via email to