[GitHub] [spark] agubichev commented on a diff in pull request #42778: [SPARK-45055] [SQL] Do not transpose windows if they conflict on ORDER BY / PROJECT clauses

["agubichev (via GitHub)" <gi...@apache.org>]

agubichev commented on code in PR #42778:
URL: https://github.com/apache/spark/pull/42778#discussion_r1315214615


##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/Optimizer.scala:
##########
@@ -1305,11 +1305,20 @@ object TransposeWindow extends Rule[LogicalPlan] {
     }
   }
 
+  // Returns false if the parent 'w1' window function order by any of the window
+  // expressions produced by 'w2'.
+  private def compatibleOrderBy(w1: Window, w2: Window): Boolean = {
+    val childWindowExprs = w2.windowExpressions.map(x => x.toAttribute)
+    val parentOrder = w1.orderSpec.flatMap(x => x.references)
+    childWindowExprs.intersect(parentOrder).isEmpty
+  }
+
   private def windowsCompatible(w1: Window, w2: Window): Boolean = {
     w1.references.intersect(w2.windowOutputSet).isEmpty &&

Review Comment:
   It's a subtle bug.
   `references` by definition do not include output attributes (see https://github.com/apache/spark/blob/798fce3b571907ee52058004cc38c2e8dbc4b016/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/plans/QueryPlan.scala#L91).
   So if the top Window function order by C and projects (outputs) C, then C is not included into references.
   Therefore, we need that extra check.
   
   (and we can't just take all attributes in the Window operator either, since PARTITION BY clauses do need to intersect)
   



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org