[CARMEL-5933] Runtime Filter supports pruning side has window (#909)

* Runtime Filter supports pruning side has window * Update DynamicDataPruningSuite.scala
apache · Apr 15, 2022 · 43597eb · 43597eb
1 parent 8c2a7dd
commit 43597eb
Show file tree

Hide file tree

Showing 2 changed files with 79 additions and 9 deletions.
diff --git a/sql/core/src/main/scala/org/apache/spark/sql/execution/dynamicpruning/PartitionPruning.scala b/sql/core/src/main/scala/org/apache/spark/sql/execution/dynamicpruning/PartitionPruning.scala
@@ -190,19 +190,33 @@ object PartitionPruning extends Rule[LogicalPlan]
     }
   }
 
+  // Make sure injected filters could push through Shuffle, see PushPredicateThroughNonJoin
+  private def probablyPushThroughShuffle(exp: Expression, plan: LogicalPlan): Boolean = {
+    plan match {
+      case j: Join if !canPlanAsBroadcastHashJoin(j, conf) => true
+      case a @ Aggregate(groupingExps, aggExps, child)
+        if aggExps.forall(_.deterministic) && groupingExps.nonEmpty &&
+          replaceAlias(exp, getAliasMap(a)).references.subsetOf(child.outputSet) => true
+      case w: Window
+        if w.partitionSpec.forall(_.isInstanceOf[AttributeReference]) &&
+          exp.references.subsetOf(AttributeSet(w.partitionSpec.flatMap(_.references))) => true
+      case p: Project =>
+        probablyPushThroughShuffle(replaceAlias(exp, getAliasMap(p)), p.child)
+      case other =>
+        other.children.exists { p =>
+          if (exp.references.subsetOf(p.outputSet)) probablyPushThroughShuffle(exp, p) else false
+        }
+    }
+  }
+
   private def dataPruningHasBenefit(
       prunRelation: LogicalRelation,
+      exp: Expression,
       prunPlan: LogicalPlan,
       otherPlan: LogicalPlan,
       canBuildBroadcast: Boolean): Boolean = {
     if (canBuildBroadcast) {
-      val shuffleStages = prunPlan.collect {
-        case j @ Join(left, right, _, _, hint)
-          if !canBroadcastBySize(left, SQLConf.get) && !canBroadcastBySize(right, SQLConf.get)
-            && !hintToBroadcastLeft(hint) && !hintToBroadcastRight(hint) => j
-        case a: Aggregate => a
-      }
-      shuffleStages.exists(_.collectLeaves().exists(_.equals(prunRelation))) &&
+      probablyPushThroughShuffle(exp, prunPlan) &&
         prunRelation.stats.sizeInBytes >= SQLConf.get.dynamicDataPruningSideThreshold
     } else {
       val estimatePruningSideSize =
@@ -251,7 +265,7 @@ object PartitionPruning extends Rule[LogicalPlan]
                 canPruneLeft(joinType) &&
                 supportDynamicPruning(right) &&
                 (canBroadcastBySize(right, conf) || hintToBroadcastRight(hint)) &&
-                dataPruningHasBenefit(scan.logicalRelation, left, right,
+                dataPruningHasBenefit(scan.logicalRelation, l, left, right,
                   canBuildBroadcastRight(joinType)) =>
                 newLeft = insertDataPredicate(l, newLeft, r, right, rightKeys)
               case _ =>
@@ -269,7 +283,7 @@ object PartitionPruning extends Rule[LogicalPlan]
                     canPruneRight(joinType) &&
                     supportDynamicPruning(left) &&
                     (canBroadcastBySize(left, conf) || hintToBroadcastLeft(hint)) &&
-                    dataPruningHasBenefit(scan.logicalRelation, right, left,
+                    dataPruningHasBenefit(scan.logicalRelation, r, right, left,
                       canBuildBroadcastLeft(joinType)) =>
                     newRight = insertDataPredicate(r, newRight, l, left, leftKeys)
                   case _ =>

diff --git a/sql/core/src/test/scala/org/apache/spark/sql/DynamicDataPruningSuite.scala b/sql/core/src/test/scala/org/apache/spark/sql/DynamicDataPruningSuite.scala
@@ -624,6 +624,62 @@ abstract class DynamicDataPruningSuiteBase
     }
   }
 
+  test("Aggregate should triggers shuffle pruning") {
+    withSQLConf(SQLConf.DYNAMIC_DATA_PRUNING_ENABLED.key -> "true",
+      SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> "3K",
+      SQLConf.DYNAMIC_DATA_PRUNING_SIDE_THRESHOLD.key -> "10K") {
+      val df = sql(
+        """
+          |SELECT t11.a,
+          |       t11.b
+          |FROM   (SELECT DISTINCT * FROM t1) t11
+          |       JOIN t3
+          |         ON t11.a = t3.a AND t3.b < 2
+          |""".stripMargin)
+
+      checkDataPruningPredicate(df, false, true)
+      checkAnswer(df, Row(0, 0) :: Row(1, 1) :: Nil)
+    }
+  }
+
+  test("Window should triggers shuffle pruning") {
+    withSQLConf(SQLConf.DYNAMIC_DATA_PRUNING_ENABLED.key -> "true",
+      SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> "3K",
+      SQLConf.DYNAMIC_DATA_PRUNING_SIDE_THRESHOLD.key -> "10K") {
+      val df = sql(
+        """
+          |SELECT t11.a,
+          |       t11.b,
+          |       t11.rn
+          |FROM   (SELECT *, Row_number() OVER (PARTITION BY a ORDER BY b) rn FROM t1) t11
+          |       JOIN t3
+          |         ON t11.a = t3.a AND t3.b < 2
+          |""".stripMargin)
+
+      checkDataPruningPredicate(df, false, true)
+      checkAnswer(df, Row(0, 0, 1) :: Row(1, 1, 1) :: Nil)
+    }
+  }
+
+  test("Window should not triggers shuffle pruning if can't push through Shuffle") {
+    withSQLConf(SQLConf.DYNAMIC_DATA_PRUNING_ENABLED.key -> "true",
+      SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> "3K",
+      SQLConf.DYNAMIC_DATA_PRUNING_SIDE_THRESHOLD.key -> "10K") {
+      val df = sql(
+        """
+          |SELECT t11.a,
+          |       t11.b,
+          |       t11.rn
+          |FROM   (SELECT *, Row_number() OVER (PARTITION BY b ORDER BY a) rn FROM t1) t11
+          |       JOIN t3
+          |         ON t11.a = t3.a AND t3.b < 2
+          |""".stripMargin)
+
+      checkDataPruningPredicate(df, false, false)
+      checkAnswer(df, Row(0, 0, 1) :: Row(1, 1, 1) :: Nil)
+    }
+  }
+
   test("CARMEL-5442: Fall back to true if InSet size exceed DYNAMIC_PRUNING_MAX_INSET_NUM") {
     withSQLConf(SQLConf.DYNAMIC_DATA_PRUNING_ENABLED.key -> "true",
       SQLConf.AUTO_BROADCASTJOIN_THRESHOLD.key -> "3K",