[
https://issues.apache.org/jira/browse/SAMZA-929?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15243372#comment-15243372
]
Chen Song commented on SAMZA-929:
---------------------------------
After some research, I believe this issue is already addressed by
https://issues.apache.org/jira/browse/YARN-2704 after 2.6.0.
[~steve_l], what you were told by your colleagues may be right. The RM will
handle both renew and recreation of HDFS delegations tokens on behalf of AM and
containers. That would address the following 2 issues.
* Localizations of failed containers and AM
* log aggregations after the job completes or dies
In order to turn on this feature, one has to do the following settings on the
hadoop cluster
* Turn on *yarn.resourcemanager.proxy-user-privileges.enabled* in yarn-site.xml
* Make yarn as a proxy user in *core-site.xml* on name nodes
* This will need a restart of name nodes and resource managers
I am going to do some testing to verify this.
> Support AM restart in Kerberos secured YARN after the original token in
> launch context expires
> ----------------------------------------------------------------------------------------------
>
> Key: SAMZA-929
> URL: https://issues.apache.org/jira/browse/SAMZA-929
> Project: Samza
> Issue Type: Sub-task
> Components: yarn
> Reporter: Yi Pan (Data Infrastructure)
> Assignee: Chen Song
>
> In Kerberos secured YARN cluster, an initial token is given in the launch
> context to RM for a job. When that token expires, RM will not be able to
> restart failed AM. As for now, we will have to re-submit the job to refresh
> the token in the launch context.
> This ticket is searching for a solution that can refresh the token in the
> launch context for AM.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
