[
https://issues.apache.org/jira/browse/HIVE-22151?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16918309#comment-16918309
]
Ashutosh Chauhan commented on HIVE-22151:
-----------------------------------------
This is a perf config. But we never got it to perform reliably. This was
introduced to prevent Mapjoin to go OOM by spilling data to disk. But what we
found in practice is if hashtable fits in memory , then this performs worse
than MapJoin and in case it doesnt and it spills perf still suffers quite a
bit. And then to determine when to spill is not easy so either you
unnecessarily spill or spill too late. Biggest issue was the first one, ie,
this impl is slow compared to mapjoin when there is no spilling.
As a result, in most sites this is turned off. Most recent instance was with
[~rameshkumar] who discovered this doesnt work very well with vectorization and
throws up exception. So, my suggestion is to have this turned off by default.
> Turn off hybrid grace hash join by default
> ------------------------------------------
>
> Key: HIVE-22151
> URL: https://issues.apache.org/jira/browse/HIVE-22151
> Project: Hive
> Issue Type: Improvement
> Reporter: Ashutosh Chauhan
> Assignee: Ashutosh Chauhan
> Priority: Major
> Attachments: HIVE-22151.patch
>
>
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
