[
https://issues.apache.org/jira/browse/HIVE-7074?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13999580#comment-13999580
]
Gopal V commented on HIVE-7074:
-------------------------------
To bring up tpc-ds 200-scale inventory table as a comparison.
{code}
hive> select inv_item_sk%200 as bucket, count(1) as total from inventory group
by inv_item_sk % 200 order by total desc limit 1;
199 225120
Time taken: 11.112 seconds, Fetched: 1 row(s)
{code}
versus 199 buckets.
{code}
hive> select inv_item_sk%199 as bucket, count(1) as total from inventory group
by inv_item_sk % 199 order by total desc limit 1;
31 190428
Time taken: 10.377 seconds, Fetched: 1 row(s)
{code}
That's 18% more for the slowest bucket while needing an extra reducer.
> The reducer parallelism should be a prime number for better stride protection
> -----------------------------------------------------------------------------
>
> Key: HIVE-7074
> URL: https://issues.apache.org/jira/browse/HIVE-7074
> Project: Hive
> Issue Type: Improvement
> Components: Statistics
> Reporter: Gopal V
> Assignee: Gopal V
> Attachments: HIVE-7074.1.patch
>
>
> The current hive reducer parallelism results in stride issues with key
> distribution.
> a JOIN generating even numbers will get strided onto only some of the
> reducers.
> The probability of distribution skew is controlled by the number of common
> factors shared by the hashcode of the key and the number of buckets.
> Using a prime number within the reducer estimation will cut that probability
> down by a significant amount.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
