[
https://issues.apache.org/jira/browse/HIVE-10704?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14563842#comment-14563842
]
Mostafa Mokhtar commented on HIVE-10704:
----------------------------------------
Table size of 0 is usually an error in the statistics and the actual table can
be arbitrary big, so evenly dividing the memory is a much safer bet.
And in the case you mentioned above each hash table will get
hive.auto.convert.join.noconditionaltask.size / 3.
By default this is Container size/3 which is more than enough for smaller
tables.
> Errors in Tez HashTableLoader when estimated table size is 0
> ------------------------------------------------------------
>
> Key: HIVE-10704
> URL: https://issues.apache.org/jira/browse/HIVE-10704
> Project: Hive
> Issue Type: Bug
> Components: Query Processor
> Reporter: Jason Dere
> Assignee: Mostafa Mokhtar
> Fix For: 1.2.1
>
> Attachments: HIVE-10704.1.patch, HIVE-10704.2.patch,
> HIVE-10704.3.patch
>
>
> Couple of issues:
> - If the table sizes in MapJoinOperator.getParentDataSizes() are 0 for all
> tables, the largest small table selection is wrong and could select the large
> table (which results in NPE)
> - The memory estimates can either divide-by-zero, or allocate 0 memory if the
> table size is 0. Try to come up with a sensible default for this.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
