[
https://issues.apache.org/jira/browse/HIVE-20254?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16578003#comment-16578003
]
Qinghui Xu commented on HIVE-20254:
-----------------------------------
I think we should backport this fix to version 1.1.0
> CheckNonCombinablePathCallable is buggy
> ---------------------------------------
>
> Key: HIVE-20254
> URL: https://issues.apache.org/jira/browse/HIVE-20254
> Project: Hive
> Issue Type: Bug
> Affects Versions: 1.1.0
> Reporter: Qinghui Xu
> Priority: Major
>
> CombineHiveInputFormat provides the possibility for people to avoid combine
> some part of their inputs (by implementing AvoidSplitCombination)
> We spot a problem with that when our query tries to read a lot of partitions
> (more than 100). In fact, when there are more than 100 input paths, the check
> of combinability is run in parallel:
> * dividing the input path array into several chunks (each chunk with no more
> than 100 paths)
> * submit each chunk to a CheckNonCombinablePathCallable
> * each CheckNonCombinablePathCallable will return a set of index for the
> paths to not be combined
> The problem is that CheckNonCombinablePathCallable returns a set of relative
> index (the index inside the chunk) instead of the absolute index, it means
> that the returned indices are always smaller than 100, thus all the paths in
> the array with position bigger than 100 are never taken into account for
> avoiding combine input.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
