hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Rui Li (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HIVE-7527) Support order by and sort by on Spark
Date Mon, 28 Jul 2014 12:36:43 GMT

    [ https://issues.apache.org/jira/browse/HIVE-7527?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14076187#comment-14076187
] 

Rui Li commented on HIVE-7527:
------------------------------

Hi [~xuefuz], I tried to run order by queries using spark's sortByKey transformation but it
seems the result is incorrect. I inserted the sortByKey between HiveMapFunction and HiveReduceFunction
(in substitute of partitionBy). Wondering if this is the right way to do it...

I detect the order by by looking at the parent ReduceSink when a ReduceWork is created and
connected to a MapWork. It worked for my simple cases :)

> Support order by and sort by on Spark
> -------------------------------------
>
>                 Key: HIVE-7527
>                 URL: https://issues.apache.org/jira/browse/HIVE-7527
>             Project: Hive
>          Issue Type: Sub-task
>          Components: Spark
>            Reporter: Xuefu Zhang
>
> Currently Hive depends completely on MapReduce's sorting as part of shuffling to achieve
order by (global sort, one reducer) and sort by (local sort).
> Spark has a sort by transformation in different variations that can used to support Hive's
order by and sort by. However, we still need to evaluate weather Spark's sortBy can achieve
the same functionality inherited from MapReduce's shuffle sort.
> Currently Hive on Spark should be able to run simple sort by or order by, by changing
the currently partitionBy to sortby. This is the way to verify theories. Complete solution
will not be available until we have complete SparkPlanGenerator.
> There is also a question of how we determine that there is order by or sort by by just
looking at the operator tree, from which Spark task is created. This is the responsibility
of SparkPlanGenerator, but we need to have an idea.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message