[
https://issues.apache.org/jira/browse/HIVE-19327?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16457112#comment-16457112
]
Sergey Shelukhin commented on HIVE-19327:
-----------------------------------------
Yeah but it may not handle all cases correctly.
The basic idea is that Hive doesn't run operators for empty splits, but for GBY
we still want to run them to generate summary (eg for rollup).
Usually Hive generates an empty split in such cases with 0 rows to force
operators to run.
This patch returns original directory of MM table if there are no valid MM
directories.
It should definitely work ok for the base case in this test - when there are no
valid MM delta directories because there's no data at all (it will be
equivalent to the custom 0-row split).
But unless I'm missing something, it won't work correctly if e.g. there are in
progress/aborted txns, so while there are no valid MM deltas, the original
directory is not empty. The split will just specify the table directory itself
and will read all these directories recursively in Tez.
> qroupby_rollup_empty.q fails for insert-only transactional tables
> -----------------------------------------------------------------
>
> Key: HIVE-19327
> URL: https://issues.apache.org/jira/browse/HIVE-19327
> Project: Hive
> Issue Type: Bug
> Components: Hive
> Affects Versions: 3.0.0
> Reporter: Steve Yeom
> Assignee: Steve Yeom
> Priority: Major
> Fix For: 3.0.0
>
> Attachments: HIVE-19327.01.patch
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
