hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Gunther Hagleitner (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HIVE-7121) Use murmur hash to distribute HiveKey
Date Tue, 27 May 2014 19:16:02 GMT

    [ https://issues.apache.org/jira/browse/HIVE-7121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14010149#comment-14010149
] 

Gunther Hagleitner commented on HIVE-7121:
------------------------------------------

[~appodictic] I think you're right. This definitely affects bucketing. 

Options I see are:

- Only do it for queries that do not enter into bucketed tables, i.e.: leave the bucketing
hash function as badly distributed as it is, but fix shuffle joins, group bys and inserts
into other tables.
- Remember the hash function in table metadata. This is slightly tricky because we probably
don't want a mix of hash functions in the same table (different partitions have different
bucketing schemes - that would probably destroy any chance of SMB on that table.) Maybe we
even want only one function per DB to make sure different tables in a DB can be joined without
looking at the hash function used for each.

How come though these unit tests are failing? I didn't think we changed the bucketing scheme
between hive 12 and 13. Did we?

> Use murmur hash to distribute HiveKey
> -------------------------------------
>
>                 Key: HIVE-7121
>                 URL: https://issues.apache.org/jira/browse/HIVE-7121
>             Project: Hive
>          Issue Type: Bug
>          Components: Query Processor
>            Reporter: Gopal V
>            Assignee: Gopal V
>         Attachments: HIVE-7121.1.patch, HIVE-7121.WIP.patch
>
>
> The current hashCode implementation produces poor parallelism when dealing with single
integers or doubles.
> And for partitioned inserts into a 1 bucket table, there is a significant hotspot on
Reducer #31.
> Removing the magic number 31 and using a more normal hash algorithm would help fix these
hotspots.



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

Mime
View raw message