[
https://issues.apache.org/jira/browse/HIVE-7782?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14170172#comment-14170172
]
Pala M Muthaia commented on HIVE-7782:
--------------------------------------
I am investigating this.
In a hive session, when the first Tez job is executed, internally the
mapreduce.framework.name property is set to "yarn-tez" (See
TezSessionState.open() method). This setting causes all native MR jobs in this
session to be sent to Tez for execution.
When the execution engine is changed to MR (i.e. set hive.execution.engine=mr),
this setting change is not reverted. So, the subsequent query, which used to go
to native MR, now goes to Tez instead (as MR on Tez). That's why its faster
than native MR, but slower than native Tez.
> tez default engine not overridden by hive.execution.engine=mr in hive cli
> session
> ---------------------------------------------------------------------------------
>
> Key: HIVE-7782
> URL: https://issues.apache.org/jira/browse/HIVE-7782
> Project: Hive
> Issue Type: Bug
> Components: CLI, Tez
> Environment: HDP2.1
> Reporter: Hari Sekhon
> Priority: Minor
> Labels: cli, hive, tez, yarn
>
> I've deployed hive.execution.engine=tez as the default on my secondary HDP
> cluster I find that hive cli interactive sessions where I do
> {code}
> set hive.execution.engine=mr
> {code}
> still execute with Tez as shown in the Resource Manager applications view.
> Now this may make sense since it's connected a Tez session by that point but
> it's also misleading because the job progress output in the cli changes to
> look like MapReduce rather than Tez and the query time is increased from 8 to
> to 15-16 secs but still less than the 25-30+ secs I usually see with MR. The
> Resource Manager shows both of these jobs as TEZ application type regardless
> of setting hive.execution.engine=mr. Is this a bug in the way Hive is
> submitting the job (Tez vs MR) or a bug in the way the RM is reporting it?
> {code}
> hive
> Logging initialized using configuration in
> file:/etc/hive/conf.dist/hive-log4j.properties
> hive> select count(*) from sample_07;
> Query ID = hari_20140819164848_c03824c7-0e76-4507-b619-6a22cb0fbc4c
> Total jobs = 1
> Launching Job 1 out of 1
> Status: Running (application id: application_1408444369445_0031)
> Map 1: -/- Reducer 2: 0/1
> Map 1: 0/1 Reducer 2: 0/1
> Map 1: 0/1 Reducer 2: 0/1
> Map 1: 1/1 Reducer 2: 0/1
> Map 1: 1/1 Reducer 2: 1/1
> Status: Finished successfully
> OK
> 823
> Time taken: 8.492 seconds, Fetched: 1 row(s)
> hive> set hive.execution.engine=mr;
> hive> select count(*) from sample_07;
> Query ID = hari_20140819164848_b620d990-b405-479c-be5b-d9616527cefe
> Total jobs = 1
> Launching Job 1 out of 1
> Number of reduce tasks determined at compile time: 1
> In order to change the average load for a reducer (in bytes):
> set hive.exec.reducers.bytes.per.reducer=<number>
> In order to limit the maximum number of reducers:
> set hive.exec.reducers.max=<number>
> In order to set a constant number of reducers:
> set mapreduce.job.reduces=<number>
> Starting Job = job_1408444369445_0032, Tracking URL =
> http://lonsl1101827-data.uk.net.intra:8088/proxy/application_1408444369445_0032/
> Kill Command = /usr/lib/hadoop/bin/hadoop job -kill job_1408444369445_0032
> Hadoop job information for Stage-1: number of mappers: 0; number of reducers: > 0
> 2014-08-19 16:48:35,242 Stage-1 map = 0%, reduce = 0%
> 2014-08-19 16:48:40,539 Stage-1 map = 100%, reduce = 0%
> 2014-08-19 16:48:44,676 Stage-1 map = 100%, reduce = 100%
> Ended Job = job_1408444369445_0032
> MapReduce Jobs Launched:
> Job 0: HDFS Read: 0 HDFS Write: 0 SUCCESS
> Total MapReduce CPU Time Spent: 0 msec
> OK
> 823
> Time taken: 16.579 seconds, Fetched: 1 row(s)
> {code}
> If I exit hive shell and restart it instead using {code}--hiveconf
> hive.execution.engine=mr{code} to set before session is established then it
> does a proper MapReduce job according to RM and it also takes the longer
> expected 25 secs instead of the 8 in Tez or 15 in trying to do MR instead Tez
> session.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
