[
https://issues.apache.org/jira/browse/HIVE-6749?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13947365#comment-13947365
]
Ashutosh Chauhan commented on HIVE-6749:
----------------------------------------
It should be turned off by default because :
* Its not clear it will always be more performant than local hashtable
generation. When this kicks in, there will be three DataNode (which has
replicas of smaller table) serving 100s of map tasks which are doing remote
reads. Compare to this, DistributedCache distributes small file as well as
localize, which will be local read. There might be some case that nonstaged is
faster, but I dont think that will be a general case.
* Seems like this will not work with security. Because we are reading smaller
table directly now from hdfs, whose tokens are not obtained.
* Its a new feature, so lets give it some time to mature.
> Turn hive.auto.convert.join.use.nonstaged off by default
> --------------------------------------------------------
>
> Key: HIVE-6749
> URL: https://issues.apache.org/jira/browse/HIVE-6749
> Project: Hive
> Issue Type: Bug
> Components: Configuration
> Affects Versions: 0.13.0
> Reporter: Ashutosh Chauhan
>
--
This message was sent by Atlassian JIRA
(v6.2#6252)
