[
https://issues.apache.org/jira/browse/HIVE-12479?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15018947#comment-15018947
]
Matt McCline commented on HIVE-12479:
-------------------------------------
Turning off the Native Vector ReduceSink does not eliminate the problem.
Current guess: old pass-thru VectorMapJoinOperator passes a separate output
batch with an identity (i.e. [0, 1, 2, 3, 4]) projection array.
VectorMapJoinMultiKeyOperator uses the input BigTable batch for output and has
a non-identity projection (i.e. [0, 3, 4, 0, 1] -- don't worry about the
repeated 0). How does the VectorFilterOperator handle projection?
> Vectorization: vectorized MultiKeyJoins with 2 integers with a filter in the
> middle
> -----------------------------------------------------------------------------------
>
> Key: HIVE-12479
> URL: https://issues.apache.org/jira/browse/HIVE-12479
> Project: Hive
> Issue Type: Bug
> Components: Vectorization
> Affects Versions: 1.3.0, 2.0.0
> Reporter: Gopal V
> Assignee: Matt McCline
> Attachments: HIVE-12479.tar.gz
>
>
> The row-counts expected with and without vectorization differ.
> The attached small-scale repro case produces 5 rows with vectorized multi-key
> joins and 53 rows without the vectorized join.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
