[
https://issues.apache.org/jira/browse/HIVE-8424?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gopal V reopened HIVE-8424:
---------------------------
This adds a hard-dependency on YARN FairScheduler jars for hive on Hadoop-2.3
shims.
{code}
import
org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler;
{code}
That class is marked clearly as
{code}
@LimitedPrivate("yarn")
@Unstable
{code}
This will fail to work entirely because the classpath is missing resource
manager JARs (all resource manager jars - capacity, fair).
With Mapreduce's rolling upgrades, YARN will only send mapreduce client JARs as
part of the classpath (which are cross-compatible).
So hive-exec.jar cannot depend on the presence of Resource Manager scheduler
JARs or any other cluster-specific JARs.
> Support fair scheduler user queue mapping in non-impersonation mode
> -------------------------------------------------------------------
>
> Key: HIVE-8424
> URL: https://issues.apache.org/jira/browse/HIVE-8424
> Project: Hive
> Issue Type: Improvement
> Components: Shims
> Reporter: Mohit Sabharwal
> Assignee: Mohit Sabharwal
> Labels: TODOC15
> Fix For: 0.15.0
>
> Attachments: HIVE-8424.1.patch, HIVE-8424.2.patch, HIVE-8424.3.patch,
> HIVE-8424.patch
>
>
> Under non-impersonation mode, all MR jobs run as the hive system user. The
> default scheduler queue mapping is one queue per user. This is problematic
> for users who use the queues to regulate and track their MR resource usage.
> Yarn exposes an API to retrieve the fair scheduler queue mapping, which we
> can use to set the appropriate MR queue for the current user.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
