[
https://issues.apache.org/jira/browse/HIVE-21465?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16811516#comment-16811516
]
Laszlo Bodor commented on HIVE-21465:
-------------------------------------
[~kgyrtkirk]: thanks for the pointer, there were some vectorization related
benchmarks, which I could have used
unfortunately bad news, with a new level in the inheritance, classes seem to
work much worse, I have to dig deeper in this ([^HIVE-21465.benchmark.patch])
{code}
Benchmark Mode Cnt Score Error Units
VectorizedExpressionBench.CastLongToDecimalBench.bench avgt 2 8504.868 ms/op
VectorizedExpressionBench.CastLongToDecimalBenchOld.bench avgt 2 4769.341 ms/op
{code}
> Introduce a base abstract VectorExpression to eliminate further code
> duplications
> ---------------------------------------------------------------------------------
>
> Key: HIVE-21465
> URL: https://issues.apache.org/jira/browse/HIVE-21465
> Project: Hive
> Issue Type: Bug
> Reporter: Laszlo Bodor
> Assignee: Laszlo Bodor
> Priority: Major
> Attachments: HIVE-21465.01.patch, HIVE-21465.02.patch,
> HIVE-21465.benchmark.patch
>
>
> In vectorized expressions, from time to time, I see a similar/same logic to
> be repeated, which always consists of some stuff like:
> 1. some inner loops (even the comments are the same)
> {code}
> grep -iRH "Set isNull before call in case it changes it mind" | grep
> "ql/src/java/org/apache/hadoop/hive/ql/exec/vector/expressions/"
> {code}
> 2. call to an actual "business logic" function ==> 1 line which could be
> abstracted... and it has been abstracted in hive codebase in n different
> cases.
> However, an abstract BaseVectorExpression would not force anybody to extend
> it instead of VectorExpression and not copy the code again, but it could be a
> good start and an example.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
