[ 
https://issues.apache.org/jira/browse/DRILL-4147?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15403078#comment-15403078
 ] 

ASF GitHub Bot commented on DRILL-4147:
---------------------------------------

Github user amansinha100 commented on a diff in the pull request:

    https://github.com/apache/drill/pull/555#discussion_r73075658
  
    --- Diff: 
exec/java-exec/src/main/java/org/apache/drill/exec/planner/physical/UnionAllPrule.java
 ---
    @@ -61,7 +61,8 @@ public void onMatch(RelOptRuleCall call) {
             convertedInputList.add(convertedInput);
           }
     
    -      traits = 
call.getPlanner().emptyTraitSet().plus(Prel.DRILL_PHYSICAL).plus(DrillDistributionTrait.SINGLETON);
    +      // distribution trait is set to ANY to allow Union-All to inherit 
the distribution of its child
    +      traits = 
call.getPlanner().emptyTraitSet().plus(Prel.DRILL_PHYSICAL).plus(DrillDistributionTrait.ANY);
    --- End diff --
    
    The union-all's distribution trait will be ANY (I should modify the comment 
since it is not accurate).   If there is another downstream operator after 
union-all, such as Aggregation or Join, that operator will impose its own 
distribution requirement.  Treating the output trait as ANY is at least better 
than the current SINGLETON but we are not propagating either the LHS or RHS 
trait...this could be an enhancement.


> Union All operator runs in a single fragment
> --------------------------------------------
>
>                 Key: DRILL-4147
>                 URL: https://issues.apache.org/jira/browse/DRILL-4147
>             Project: Apache Drill
>          Issue Type: Bug
>            Reporter: amit hadke
>            Assignee: Aman Sinha
>
> A User noticed that running select  from a single directory is much faster 
> than union all on two directories.
> (https://drill.apache.org/blog/2014/12/09/running-sql-queries-on-amazon-s3/#comment-2349732267)
>  
> It seems like UNION ALL operator doesn't parallelize sub scans (its using 
> SINGLETON for distribution type). Everything is ran in single fragment.
> We may have to use SubsetTransformer in UnionAllPrule.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to