hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chengxiang Li (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (HIVE-7675) Implement native HiveMapFunction
Date Mon, 11 Aug 2014 06:47:11 GMT

     [ https://issues.apache.org/jira/browse/HIVE-7675?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Chengxiang Li updated HIVE-7675:
--------------------------------

    Description: 
Currently, Hive on Spark depend on ExecMapper to execute operator logic, full stack is like:
Spark FrameWork=>HiveMapFunction=>ExecMapper=>Hive operators. HiveMapFunction is
just a thin wrapper of ExecMapper, this introduce several problems as following:
# ExecMapper is designed for MR single process task mode, it does not work well under Spark
multi-thread task node.
# ExecMapper introduce extra API level restriction and process logic.

We need implement native HiveMapFunction, as the bridge between Spark framework and Hive operators.

  was:
Currently, Hive on Spark depend on ExecMapper to execute operator logic, full stack is like:
Spark FrameWork=>HiveMapFunction=>ExecMapper=>Hive operators. HiveMapFunction is
just a thin wrapper of ExecMapper, this introduce several problems as following:
# ExecMapper is designed for MR single process task mode, it does not work well under Spark
multi-thread task node.
# ExecMapper introduce extra API level restriction.

We need implement native HiveMapFunction, as the bridge between Spark framework and Hive operators.


> Implement native HiveMapFunction
> --------------------------------
>
>                 Key: HIVE-7675
>                 URL: https://issues.apache.org/jira/browse/HIVE-7675
>             Project: Hive
>          Issue Type: Sub-task
>          Components: Spark
>            Reporter: Chengxiang Li
>
> Currently, Hive on Spark depend on ExecMapper to execute operator logic, full stack is
like: Spark FrameWork=>HiveMapFunction=>ExecMapper=>Hive operators. HiveMapFunction
is just a thin wrapper of ExecMapper, this introduce several problems as following:
> # ExecMapper is designed for MR single process task mode, it does not work well under
Spark multi-thread task node.
> # ExecMapper introduce extra API level restriction and process logic.
> We need implement native HiveMapFunction, as the bridge between Spark framework and Hive
operators.



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

Mime
View raw message