Tao Jie commented on YARN-5636:

We did some trial like this:
# Add fields to ResourceRequest, ContainerRequest:
#* String reservedLabel,  to recognize those containers that resources reserved 
to (actually is slider app name)
#* boolean useReserved, tell the RM that whether this container needs to use 
reserved resource (true when it is a restarted container)
# When completing container on node, it will mark that resource as *reserved  
resource* if this container has a reservedLabel
# When assigning resource to container on node, it will check available 
resource on node respect to reserved resource.
# Also reserved resource would release when it expires.
[~Naganarasimha], [~sunilg], [~leftnoteasy], Like to hear your thoughts.  

> Support reserving resources on certain nodes for certain applications
> ---------------------------------------------------------------------
>                 Key: YARN-5636
>                 URL: https://issues.apache.org/jira/browse/YARN-5636
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: scheduler
>            Reporter: Tao Jie
> We have met such circumstance:
> We are trying to run storm&kafka on yarn by Slider, and Storm&Kafka writes 
> data to local disk on node. If some containers or the application fails, we 
> expect that those containers would restart on the same node as they run 
> before, otherwise data written on local would lost.
> For slider, it will trying to ensure restarted container on same nodes as 
> before. However for yarn, resource may be assigned to other applications when 
> former long-running application is down.
> As a result we'd better to have a mechanism that reserve some resource for 
> certain long-running applications on certain nodes for a period of time. Does 
> it make sense?

This message was sent by Atlassian JIRA

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