maropu commented on a change in pull request #27058: [SPARK-30276][SQL] Support
Filter expression allows simultaneous use of DISTINCT
URL: https://github.com/apache/spark/pull/27058#discussion_r367724105
##########
File path:
sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/optimizer/RewriteDistinctAggregates.scala
##########
@@ -151,21 +197,101 @@ object RewriteDistinctAggregates extends
Rule[LogicalPlan] {
// We need at least two distinct aggregates for this rule because
aggregation
// strategy can handle a single distinct group.
// This check can produce false-positives, e.g., SUM(DISTINCT a) &
COUNT(DISTINCT a).
- distinctAggs.size > 1
+ distinctAggs.size >= 1
}
def apply(plan: LogicalPlan): LogicalPlan = plan transformUp {
case a: Aggregate if mayNeedtoRewrite(a.aggregateExpressions) => rewrite(a)
}
def rewrite(a: Aggregate): Aggregate = {
+ val expandAggregate = expandDistinctAggregateWithFilter(a)
+ rewriteDistinctAggregate(expandAggregate)
+ }
- // Collect all aggregate expressions.
- val aggExpressions = a.aggregateExpressions.flatMap { e =>
- e.collect {
- case ae: AggregateExpression => ae
+ private def expandDistinctAggregateWithFilter(a: Aggregate): Aggregate = {
+ val aggExpressions = collectAggregateExprs(a)
+ val (distinctAggExpressions, regularAggExpressions) =
aggExpressions.partition(_.isDistinct)
+ if (distinctAggExpressions.exists(_.filter.isDefined)) {
+ // Setup expand for the 'regular' aggregate expressions.
+ val regularAggExprs = regularAggExpressions.filter(e =>
e.children.exists(!_.foldable))
+ val regularFunChildren = regularAggExprs
+ .flatMap(_.aggregateFunction.children.filter(!_.foldable))
+ val regularFilterAttrs = regularAggExprs.flatMap(_.filterAttributes)
+ val regularAggChildren = (regularFunChildren ++
regularFilterAttrs).distinct
+ val regularAggChildAttrMap =
regularAggChildren.map(expressionAttributePair)
+ val regularAggChildAttrLookup = regularAggChildAttrMap.toMap
+ val regularOperatorMap = regularAggExprs.map {
+ case ae @ AggregateExpression(af, _, _, filter, _) =>
+ val newChildren = af.children.map(c =>
regularAggChildAttrLookup.getOrElse(c, c))
+ val raf =
af.withNewChildren(newChildren).asInstanceOf[AggregateFunction]
+ val filterOpt = filter.map(_.transform {
+ case a: Attribute => regularAggChildAttrLookup.getOrElse(a, a)
+ })
+ val aggExpr = ae.copy(aggregateFunction = raf, filter = filterOpt)
+ (ae, aggExpr)
+ }
+
+ // Setup expand for the distinct aggregate expressions.
+ val distinctAggExprs = distinctAggExpressions.filter(e =>
e.children.exists(!_.foldable))
+ val rewriteDistinctOperatorMap = distinctAggExprs.map {
+ case ae @ AggregateExpression(af, _, _, filter, _) =>
+ // Why do we need to construct the phantom id ?
+ // First, In order to reduce costs, it is better to handle the
filter clause locally.
+ // e.g. COUNT (DISTINCT a) FILTER (WHERE id > 1), evaluate expression
+ // If(id > 1) 'a else null first, and use the result as output.
+ // Second, If more than one DISTINCT aggregate expression uses the
same column,
+ // We need to construct the phantom attributes so as the output not
lost.
+ // e.g. SUM (DISTINCT a), COUNT (DISTINCT a) FILTER (WHERE id > 1)
will output
+ // attribute '_gen_distinct-1 and attribute '_gen_distinct-2 instead
of two 'a.
+ // Note: We just need to illusion the expression with filter clause.
+ // The illusionary mechanism may result in multiple distinct
aggregations uses
+ // different column, so we still need to call `rewrite`.
+ val phantomId = NamedExpression.newExprId.id
+ val unfoldableChildren = af.children.filter(!_.foldable)
+ val exprAttrs = unfoldableChildren.map { e =>
+ (e, AttributeReference(s"_gen_distinct_$phantomId", e.dataType,
nullable = true)())
Review comment:
nit: `phantomId` -> `exprId`
----------------------------------------------------------------
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]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
