fix: Comet columnar shuffle should not be on top of another Comet shu…

…ffle operator (apache#296) * fix: Comet columnar shuffle should not be on top of another Comet shuffle operator * Fix 3.2 build
himadripal · Apr 22, 2024 · adb3682 · adb3682
1 parent f97392d
commit adb3682
Show file tree

Hide file tree

Showing 2 changed files with 43 additions and 4 deletions.
diff --git a/spark/src/main/scala/org/apache/comet/CometSparkSessionExtensions.scala b/spark/src/main/scala/org/apache/comet/CometSparkSessionExtensions.scala
@@ -151,7 +151,8 @@ class CometSparkSessionExtensions
         case s: ShuffleExchangeExec
             if (!s.child.supportsColumnar || isCometPlan(
               s.child)) && isCometColumnarShuffleEnabled(conf) &&
-              QueryPlanSerde.supportPartitioningTypes(s.child.output) =>
+              QueryPlanSerde.supportPartitioningTypes(s.child.output) &&
+              !isShuffleOperator(s.child) =>
           logInfo("Comet extension enabled for JVM Columnar Shuffle")
           CometShuffleExchangeExec(s, shuffleType = CometColumnarShuffle)
       }
@@ -538,10 +539,13 @@ class CometSparkSessionExtensions
           }
 
         // Columnar shuffle for regular Spark operators (not Comet) and Comet operators
-        // (if configured)
+        // (if configured).
+        // If the child of ShuffleExchangeExec is also a ShuffleExchangeExec, we should not
+        // convert it to CometColumnarShuffle,
         case s: ShuffleExchangeExec
             if isCometShuffleEnabled(conf) && isCometColumnarShuffleEnabled(conf) &&
-              QueryPlanSerde.supportPartitioningTypes(s.child.output) =>
+              QueryPlanSerde.supportPartitioningTypes(s.child.output) &&
+              !isShuffleOperator(s.child) =>
           logInfo("Comet extension enabled for JVM Columnar Shuffle")
 
           val newOp = QueryPlanSerde.operator2Proto(s)
@@ -627,6 +631,18 @@ class CometSparkSessionExtensions
         case s: ShuffleQueryStageExec => findPartialAgg(s.plan)
       }.flatten
     }
+
+    /**
+     * Returns true if a given spark plan is Comet shuffle operator.
+     */
+    def isShuffleOperator(op: SparkPlan): Boolean = {
+      op match {
+        case op: ShuffleQueryStageExec if op.plan.isInstanceOf[CometShuffleExchangeExec] => true
+        case _: CometShuffleExchangeExec => true
+        case op: CometSinkPlaceHolder => isShuffleOperator(op.child)
+        case _ => false
+      }
+    }
   }
 
   // This rule is responsible for eliminating redundant transitions between row-based and

diff --git a/spark/src/test/scala/org/apache/comet/exec/CometExecSuite.scala b/spark/src/test/scala/org/apache/comet/exec/CometExecSuite.scala
@@ -40,7 +40,8 @@ import org.apache.spark.sql.execution.{CollectLimitExec, ProjectExec, SQLExecuti
 import org.apache.spark.sql.execution.exchange.BroadcastExchangeExec
 import org.apache.spark.sql.execution.joins.{BroadcastNestedLoopJoinExec, CartesianProductExec, SortMergeJoinExec}
 import org.apache.spark.sql.execution.window.WindowExec
-import org.apache.spark.sql.functions.{col, date_add, expr, sum}
+import org.apache.spark.sql.expressions.Window
+import org.apache.spark.sql.functions.{col, date_add, expr, lead, sum}
 import org.apache.spark.sql.internal.SQLConf
 import org.apache.spark.sql.internal.SQLConf.SESSION_LOCAL_TIMEZONE
 import org.apache.spark.unsafe.types.UTF8String
@@ -60,6 +61,28 @@ class CometExecSuite extends CometTestBase {
     }
   }
 
+  test("Repeated shuffle exchange don't fail") {
+    assume(isSpark33Plus)
+    Seq("true", "false").foreach { aqeEnabled =>
+      withSQLConf(
+        SQLConf.ADAPTIVE_EXECUTION_ENABLED.key -> aqeEnabled,
+        // `REQUIRE_ALL_CLUSTER_KEYS_FOR_DISTRIBUTION` is a new config in Spark 3.3+.
+        "spark.sql.requireAllClusterKeysForDistribution" -> "true",
+        CometConf.COMET_COLUMNAR_SHUFFLE_ENABLED.key -> "true") {
+        val df =
+          Seq(("a", 1, 1), ("a", 2, 2), ("b", 1, 3), ("b", 1, 4)).toDF("key1", "key2", "value")
+        val windowSpec = Window.partitionBy("key1", "key2").orderBy("value")
+
+        val windowed = df
+          // repartition by subset of window partitionBy keys which satisfies ClusteredDistribution
+          .repartition($"key1")
+          .select(lead($"key1", 1).over(windowSpec), lead($"value", 1).over(windowSpec))
+
+        checkSparkAnswer(windowed)
+      }
+    }
+  }
+
   test("try_sum should return null if overflow happens before merging") {
     assume(isSpark33Plus, "try_sum is available in Spark 3.3+")
     val longDf = Seq(Long.MaxValue, Long.MaxValue, 2).toDF("v")