[jira] [Comment Edited] (HIVE-13617) LLAP: support non-vectorized execution in IO

["Sergey Shelukhin (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-13617?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15280951#comment-15280951 ] 

Sergey Shelukhin edited comment on HIVE-13617 at 5/12/16 12:21 AM:
-------------------------------------------------------------------

Preliminary patch. I need to double check that results from simple types are correct.
Also complex types do not appear to work in the elevator, would need to double check that too.

I gave up on trying to make a universal converter because it's nigh impossible to substitute the OIs... attaching the terminated-in-progress patch for that too, in case it's ever needed.


was (Author: sershe):
Preliminary patch. I need to double check that results from simple types are correct.
Also complex types do not appear to work in the elevator, would need to double check that too.

I gave up on trying to make a universal converted because it's nigh impossible to substitute the OIs... attaching the terminated-in-progress patch for that too, in case it's ever needed.

> LLAP: support non-vectorized execution in IO
> --------------------------------------------
>
>                 Key: HIVE-13617
>                 URL: https://issues.apache.org/jira/browse/HIVE-13617
>             Project: Hive
>          Issue Type: Bug
>            Reporter: Sergey Shelukhin
>            Assignee: Sergey Shelukhin
>         Attachments: HIVE-13617-wo-11417.patch, HIVE-13617.patch, HIVE-15396-with-oi.patch
>
>
> Two approaches - a separate decoding path, into rows instead of VRBs; or decoding VRBs into rows on a higher level (the original LlapInputFormat). I think the latter might be better - it's not a hugely important path, and perf in non-vectorized case is not the best anyway, so it's better to make do with much less new code and architectural disruption. 
> Some ORC patches in progress introduce an easy to reuse (or so I hope, anyway) VRB-to-row conversion, so we should just use that.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)