[
https://issues.apache.org/jira/browse/SAMZA-925?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15435473#comment-15435473
]
Navina Ramesh commented on SAMZA-925:
-------------------------------------
Does the code for this functionality belong in the ConfigManager within
samza-autoscaling project as may previous patch was?
>> No. If this functionality were in ConfigManager, then it would be external
>> to the Samza framework itself.
Should this be the component that polls the JobModel and then invokes asks the
application master to restart the containers when there has been a change in
the JobModel?
>> Yes. We want the component that polls for JobModel updates to also handle
>> the restart of containers. The design would probably look very similar what
>> has been proposed in SAMZA-980.
Which APIs should be used for actually killing/starting the containers for a
particular application? Perhaps using
org.apache.hadoop.yarn.client.api.AMRMClient?
>> I think you should be using the NMClient APIs to start/stop the containers.
>> You can use AMRMClient when you want to re-allocate the containers.
> Explicit restart containers to pick up dynamic JobModel changes
> ---------------------------------------------------------------
>
> Key: SAMZA-925
> URL: https://issues.apache.org/jira/browse/SAMZA-925
> Project: Samza
> Issue Type: Sub-task
> Reporter: Yi Pan (Data Infrastructure)
> Assignee: Alex Buck
>
> When JobModel changes, the Job Coordinator should trigger the AM to kill the
> existing containers and spawn new ones based on the JobModel update. This
> way, the application remains alive and only the partitions get re-distributed
> among the containers.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
