Ngone51 commented on a change in pull request #28869:
URL: https://github.com/apache/spark/pull/28869#discussion_r443302172
##########
File path:
sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/BaseAggregateExec.scala
##########
@@ -53,15 +53,31 @@ trait BaseAggregateExec extends UnaryExecNode {
// can't bind the `mergeExpressions` with the output of the partial
aggregate, as they use
// the `inputAggBufferAttributes` of the original `DeclarativeAggregate`
before copy. Instead,
// we shall use `inputAggBufferAttributes` after copy to match the new
`mergeExpressions`.
- val aggAttrs = aggregateExpressions
- // there're exactly four cases needs `inputAggBufferAttributes` from
child according to the
- // agg planning in `AggUtils`: Partial -> Final, PartialMerge -> Final,
- // Partial -> PartialMerge, PartialMerge -> PartialMerge.
- .filter(a => a.mode == Final || a.mode ==
PartialMerge).map(_.aggregateFunction)
- .flatMap(_.inputAggBufferAttributes)
+ val aggAttrs = inputAggBufferAttributes
child.output.dropRight(aggAttrs.length) ++ aggAttrs
} else {
child.output
}
}
+
+ private val inputAggBufferAttributes: Seq[Attribute] = {
+ aggregateExpressions
+ // there're exactly four cases needs `inputAggBufferAttributes` from
child according to the
+ // agg planning in `AggUtils`: Partial -> Final, PartialMerge -> Final,
+ // Partial -> PartialMerge, PartialMerge -> PartialMerge.
+ .filter(a => a.mode == Final || a.mode == PartialMerge)
+ .flatMap(_.aggregateFunction.inputAggBufferAttributes)
+ }
+
+ protected val aggregateBufferAttributes: Seq[AttributeReference] = {
+ aggregateExpressions.flatMap(_.aggregateFunction.aggBufferAttributes)
+ }
+
+ override def producedAttributes: AttributeSet =
+ AttributeSet(aggregateAttributes) ++
+
AttributeSet(resultExpressions.diff(groupingExpressions).map(_.toAttribute)) ++
+ AttributeSet(aggregateBufferAttributes) ++
+ // it's not empty when the inputAggBufferAttributes is from the child
Aggregate, which contains
+ // subquery in AggregateFunction. See SPARK-31620 for more details.
+ AttributeSet(inputAggBufferAttributes.filterNot(child.output.contains))
Review comment:
Oh..the comment should actually be:
`it's not empty when the child Aggregate contains the subquery in
AggregateFunction.`
After SPARK-31620, the inputAggBufferAttributes is not from children but the
node itself when there're Final/PartialMerge aggregate expressions, see
`inputAttributes()` above. (But please note that agg attributes are still the
same between the parent agg node and child agg node when there's no subquery in
agg expression)
Therefore, in the case of SPARK-31620, we actually use the attributes
produced by the node itself. But for other cases, we still use the agg buffer
attributes from the children, since `inputAggBufferAttributes` is equal to the
agg buffer attributes from the children and so
`inputAggBufferAttributes.filterNot(child.output.contains)` is empty.
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
