uros-b commented on code in PR #57629:
URL: https://github.com/apache/spark/pull/57629#discussion_r3700556838
##########
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/InsertMapSortExpression.scala:
##########
@@ -52,32 +65,103 @@ object InsertMapSortInGroupingExpressions extends
Rule[LogicalPlan] {
plan transformUpWithNewOutput {
case agg @ Aggregate(groupingExprs, aggregateExpressions, child, hint) =>
- val exprToMapSort = new mutable.HashMap[Expression, NamedExpression]
- val newGroupingKeys = groupingExprs.map { expr =>
- val inserted = insertMapSortRecursively(expr)
- if (expr.ne(inserted)) {
- exprToMapSort.getOrElseUpdate(
- expr.canonicalized,
- Alias(inserted, "_groupingmapsort")()
- ).toAttribute
+ val distinctExpressions = if (normalizeDistinctAggregates) {
+ distinctAggregateChildren(aggregateExpressions)
+ } else {
+ Seq.empty
+ }
+ val expressionsToNormalize = groupingExprs ++ distinctExpressions
+ if (!expressionsToNormalize.exists(mapTypeExistsRecursively)) {
+ agg -> Nil
+ } else {
+ val groupingMapSortAliases = insertMapSortInExpressions(
+ expressions = groupingExprs,
+ aliasName = "_groupingmapsort")
+ val distinctInputAliases =
createDistinctInputAliases(distinctExpressions)
+ val distinctMapSortAliases = insertMapSortInExpressions(
+ expressions = distinctExpressions,
+ aliasName = "_distinctmapsort",
+ reusableAliases = groupingMapSortAliases,
+ inputAliases = distinctInputAliases)
+ val newGroupingKeys = groupingExprs.map { expr =>
+
groupingMapSortAliases.get(expr.canonicalized).map(_.toAttribute).getOrElse(expr)
+ }
+ val newAggregateExprs = aggregateExpressions.map {
+ case named if groupingMapSortAliases.contains(named.canonicalized)
=>
+ // If we replace the top-level named expr, then should add back
the original name
+
groupingMapSortAliases(named.canonicalized).toAttribute.withName(named.name)
+ case other =>
+ // This must be top-down so distinct arguments are normalized
before their children.
+ // Continuing downward also substitutes grouping aliases in
other arguments.
+ other.transformDown {
+ case ae: AggregateExpression if ae.isDistinct =>
+ ae.copy(aggregateFunction =
ae.aggregateFunction.withNewChildren(
+ ae.aggregateFunction.children.map { child =>
+ distinctMapSortAliases.get(child.canonicalized)
+ .map(_.toAttribute).getOrElse(child)
+ }).asInstanceOf[AggregateFunction])
+ case e =>
+
groupingMapSortAliases.get(e.canonicalized).map(_.toAttribute).getOrElse(e)
+ }.asInstanceOf[NamedExpression]
+ }
+ val distinctInput = if (distinctInputAliases.nonEmpty) {
+ // Project complex inputs once because recursive normalization can
reference them
+ // repeatedly.
Review Comment:
Nit: Could we extend this comment to say the extra layers don't survive?
FinishAnalysis runs before the operator optimization batch, so CollapseProject
and ColumnPruning fold both projections away. Reading the rule in isolation,
two stacked Projects over the aggregate's child look like a plan-size
regression, and a future reader shouldn't have to reconstruct the batch
ordering to convince themselves otherwise.
--
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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
