[
https://issues.apache.org/jira/browse/DRILL-351?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13885634#comment-13885634
]
Julian Hyde commented on DRILL-351:
-----------------------------------
Most projections can (and should) be pushed down below the join. The join keys
are the only columns that have to be present, and they are usually small
(integers).
So, marginal benefit if you are pushing down projections, as you should be.
> MergeJoin physical operator projects columns that may not be needed
> -------------------------------------------------------------------
>
> Key: DRILL-351
> URL: https://issues.apache.org/jira/browse/DRILL-351
> Project: Apache Drill
> Issue Type: Bug
> Reporter: Aman Sinha
>
> The MergeJoin physical operator currently projects all columns from the left
> and right inputs with the assumption that a subsequent Project will drop the
> unnecessary columns. This is sub-optimal because we incur the memory and cpu
> overhead of populating the output record batch value vectors for those
> columns. Ideally, the operator could keep track of the columns that are
> needed by the parent (downstream) operator and only project those columns.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
