[ 
https://issues.apache.org/jira/browse/HIVE-17405?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16150009#comment-16150009
 ] 

Sahil Takiar commented on HIVE-17405:
-------------------------------------

[~lirui] updated patch with better comments. I think the call to constant 
propagation was probably originally added just for DPP, but there are benefits 
to running it after all the operator optimizations have run in the 
{{SparkCompiler}}. Some of the optimizations (such as DPP) may have modified 
the operator tree such that additional constants can be propagated.

The modifications to the skew join qtests are one example.

> HoS DPP ConstantPropagate should use ConstantPropagateOption.SHORTCUT
> ---------------------------------------------------------------------
>
>                 Key: HIVE-17405
>                 URL: https://issues.apache.org/jira/browse/HIVE-17405
>             Project: Hive
>          Issue Type: Sub-task
>          Components: Spark
>            Reporter: Sahil Takiar
>            Assignee: Sahil Takiar
>         Attachments: HIVE-17405.1.patch, HIVE-17405.2.patch, 
> HIVE-17405.3.patch, HIVE-17405.4.patch, HIVE-17405.5.patch, 
> HIVE-17405.6.patch, HIVE-17405.7.patch
>
>
> In {{SparkCompiler#runDynamicPartitionPruning}} we should change {{new 
> ConstantPropagate().transform(parseContext)}} to {{new 
> ConstantPropagate(ConstantPropagateOption.SHORTCUT).transform(parseContext)}}
> Hive-on-Tez does the same thing.
> Running the full constant propagation isn't really necessary, we just want to 
> eliminate any {{and true}} predicates that were introduced by 
> {{SyntheticJoinPredicate}} and {{DynamicPartitionPruningOptimization}}. The 
> {{SyntheticJoinPredicate}} will introduce dummy filter predicates into the 
> operator tree, and {{DynamicPartitionPruningOptimization}} will replace them. 
> The predicates introduced via {{SyntheticJoinPredicate}} are necessary to 
> help {{DynamicPartitionPruningOptimization}} determine if DPP can be used or 
> not.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Reply via email to