[SPARK-51008][SQL] Add ResultStage for AQE

sql/catalyst/src/main/scala/org/apache/spark/sql/internal/StaticSQLConf.scala

@@ -210,6 +210,15 @@ object StaticSQLConf {
       .checkValue(thres => thres > 0 && thres <= 128, "The threshold must be in (0,128].")
       .createWithDefault(16)
 
+  val RESULT_QUERY_STAGE_MAX_THREAD_THRESHOLD =
+    buildStaticConf("spark.sql.resultQueryStage.maxThreadThreshold")
+      .internal()
+      .doc("The maximum degree of parallelism to execute ResultQueryStageExec in AQE")
+      .version("4.0.0")
+      .intConf
+      .checkValue(thres => thres > 0 && thres <= 1024, "The threshold must be in (0,1024].")
+      .createWithDefault(1024)
+
   val SQL_EVENT_TRUNCATE_LENGTH = buildStaticConf("spark.sql.event.truncate.length")
     .doc("Threshold of SQL length beyond which it will be truncated before adding to " +
       "event. Defaults to no truncation. If set to 0, callsite will be logged instead.")

sql/core/src/main/scala/org/apache/spark/sql/execution/SQLExecution.scala

@@ -17,7 +17,7 @@
 
 package org.apache.spark.sql.execution
 
-import java.util.concurrent.{ConcurrentHashMap, ExecutorService, Future => JFuture}
+import java.util.concurrent.{CompletableFuture, ConcurrentHashMap, ExecutorService}
 import java.util.concurrent.atomic.AtomicLong
 
 import scala.jdk.CollectionConverters._
@@ -301,15 +301,15 @@ object SQLExecution extends Logging {
    * SparkContext local properties are forwarded to execution thread
    */
   def withThreadLocalCaptured[T](
-      sparkSession: SparkSession, exec: ExecutorService) (body: => T): JFuture[T] = {
+      sparkSession: SparkSession, exec: ExecutorService) (body: => T): CompletableFuture[T] = {
     val activeSession = sparkSession
     val sc = sparkSession.sparkContext
     val localProps = Utils.cloneProperties(sc.getLocalProperties)
     // `getCurrentJobArtifactState` will return a stat only in Spark Connect mode. In non-Connect
     // mode, we default back to the resources of the current Spark session.
     val artifactState = JobArtifactSet.getCurrentJobArtifactState.getOrElse(
       activeSession.artifactManager.state)
-    exec.submit(() => JobArtifactSet.withActiveJobArtifactState(artifactState) {
+    CompletableFuture.supplyAsync(() => JobArtifactSet.withActiveJobArtifactState(artifactState) {
       val originalSession = SparkSession.getActiveSession
       val originalLocalProps = sc.getLocalProperties
       SparkSession.setActiveSession(activeSession)
@@ -326,6 +326,6 @@ object SQLExecution extends Logging {
         SparkSession.clearActiveSession()
       }
       res
-    })
+    }, exec)
   }
 }

sql/core/src/main/scala/org/apache/spark/sql/execution/adaptive/AdaptiveSparkPlanExec.scala

@@ -268,9 +268,11 @@ case class AdaptiveSparkPlanExec(
 
   def finalPhysicalPlan: SparkPlan = withFinalPlanUpdate(identity)
 
-  private def getFinalPhysicalPlan(): SparkPlan = lock.synchronized {
-    if (isFinalPlan) return currentPhysicalPlan
-
+  /**
+   * Run `fun` on finalized physical plan
+   */
+  def withFinalPlanUpdate[T](fun: SparkPlan => T): T = lock.synchronized {
+    _isFinalPlan = false
     // In case of this adaptive plan being executed out of `withActive` scoped functions, e.g.,
     // `plan.queryExecution.rdd`, we need to set active session here as new plan nodes can be
     // created in the middle of the execution.
@@ -279,7 +281,7 @@ case class AdaptiveSparkPlanExec(
       // Use inputPlan logicalLink here in case some top level physical nodes may be removed
       // during `initialPlan`
       var currentLogicalPlan = inputPlan.logicalLink.get
-      var result = createQueryStages(currentPhysicalPlan)
+      var result = createQueryStages(fun, currentPhysicalPlan, firstRun = true)
       val events = new LinkedBlockingQueue[StageMaterializationEvent]()
       val errors = new mutable.ArrayBuffer[Throwable]()
       var stagesToReplace = Seq.empty[QueryStageExec]
@@ -344,56 +346,53 @@ case class AdaptiveSparkPlanExec(
         if (errors.nonEmpty) {
           cleanUpAndThrowException(errors.toSeq, None)
         }
-
-        // Try re-optimizing and re-planning. Adopt the new plan if its cost is equal to or less
-        // than that of the current plan; otherwise keep the current physical plan together with
-        // the current logical plan since the physical plan's logical links point to the logical
-        // plan it has originated from.
-        // Meanwhile, we keep a list of the query stages that have been created since last plan
-        // update, which stands for the "semantic gap" between the current logical and physical
-        // plans. And each time before re-planning, we replace the corresponding nodes in the
-        // current logical plan with logical query stages to make it semantically in sync with
-        // the current physical plan. Once a new plan is adopted and both logical and physical
-        // plans are updated, we can clear the query stage list because at this point the two plans
-        // are semantically and physically in sync again.
-        val logicalPlan = replaceWithQueryStagesInLogicalPlan(currentLogicalPlan, stagesToReplace)
-        val afterReOptimize = reOptimize(logicalPlan)
-        if (afterReOptimize.isDefined) {
-          val (newPhysicalPlan, newLogicalPlan) = afterReOptimize.get
-          val origCost = costEvaluator.evaluateCost(currentPhysicalPlan)
-          val newCost = costEvaluator.evaluateCost(newPhysicalPlan)
-          if (newCost < origCost ||
-            (newCost == origCost && currentPhysicalPlan != newPhysicalPlan)) {
-            lazy val plans =
-              sideBySide(currentPhysicalPlan.treeString, newPhysicalPlan.treeString).mkString("\n")
-            logOnLevel(log"Plan changed:\n${MDC(QUERY_PLAN, plans)}")
-            cleanUpTempTags(newPhysicalPlan)
-            currentPhysicalPlan = newPhysicalPlan
-            currentLogicalPlan = newLogicalPlan
-            stagesToReplace = Seq.empty[QueryStageExec]
+        if (!currentPhysicalPlan.isInstanceOf[ResultQueryStageExec]) {
+          // Try re-optimizing and re-planning. Adopt the new plan if its cost is equal to or less
+          // than that of the current plan; otherwise keep the current physical plan together with
+          // the current logical plan since the physical plan's logical links point to the logical
+          // plan it has originated from.
+          // Meanwhile, we keep a list of the query stages that have been created since last plan
+          // update, which stands for the "semantic gap" between the current logical and physical
+          // plans. And each time before re-planning, we replace the corresponding nodes in the
+          // current logical plan with logical query stages to make it semantically in sync with
+          // the current physical plan. Once a new plan is adopted and both logical and physical
+          // plans are updated, we can clear the query stage list because at this point the two
+          // plans are semantically and physically in sync again.
+          val logicalPlan = replaceWithQueryStagesInLogicalPlan(currentLogicalPlan, stagesToReplace)
+          val afterReOptimize = reOptimize(logicalPlan)
+          if (afterReOptimize.isDefined) {
+            val (newPhysicalPlan, newLogicalPlan) = afterReOptimize.get
+            val origCost = costEvaluator.evaluateCost(currentPhysicalPlan)
+            val newCost = costEvaluator.evaluateCost(newPhysicalPlan)
+            if (newCost < origCost ||
+              (newCost == origCost && currentPhysicalPlan != newPhysicalPlan)) {
+              lazy val plans = sideBySide(
+                currentPhysicalPlan.treeString, newPhysicalPlan.treeString).mkString("\n")
+              logOnLevel(log"Plan changed:\n${MDC(QUERY_PLAN, plans)}")
+              cleanUpTempTags(newPhysicalPlan)
+              currentPhysicalPlan = newPhysicalPlan
+              currentLogicalPlan = newLogicalPlan
+              stagesToReplace = Seq.empty[QueryStageExec]
+            }
           }
         }
         // Now that some stages have finished, we can try creating new stages.
-        result = createQueryStages(currentPhysicalPlan)
+        result = createQueryStages(fun, currentPhysicalPlan, firstRun = false)
       }
-
-      // Run the final plan when there's no more unfinished stages.
-      currentPhysicalPlan = applyPhysicalRules(
-        optimizeQueryStage(result.newPlan, isFinalStage = true),
-        postStageCreationRules(supportsColumnar),
-        Some((planChangeLogger, "AQE Post Stage Creation")))
-      _isFinalPlan = true
-      executionId.foreach(onUpdatePlan(_, Seq(currentPhysicalPlan)))
-      currentPhysicalPlan
     }
+    _isFinalPlan = true
+    finalPlanUpdate
+    // Dereference the result so it can be GCed. After this resultStage.isMaterialized will return
+    // false, which is expected. If we want to collect result again, we should invoke
+    // `withFinalPlanUpdate` and pass another result handler and we will create a new result stage.
+    currentPhysicalPlan.asInstanceOf[ResultQueryStageExec].resultOption.getAndUpdate(_ => None)
+      .get.asInstanceOf[T]
   }
 
   // Use a lazy val to avoid this being called more than once.
   @transient private lazy val finalPlanUpdate: Unit = {
-    // Subqueries that don't belong to any query stage of the main query will execute after the
-    // last UI update in `getFinalPhysicalPlan`, so we need to update UI here again to make sure
-    // the newly generated nodes of those subqueries are updated.
-    if (shouldUpdatePlan && currentPhysicalPlan.exists(_.subqueries.nonEmpty)) {
+    // Do final plan update after result stage has materialized.
+    if (shouldUpdatePlan) {
       getExecutionId.foreach(onUpdatePlan(_, Seq.empty))
     }
     logOnLevel(log"Final plan:\n${MDC(QUERY_PLAN, currentPhysicalPlan)}")
@@ -426,13 +425,6 @@ case class AdaptiveSparkPlanExec(
     }
   }
 
-  private def withFinalPlanUpdate[T](fun: SparkPlan => T): T = {
-    val plan = getFinalPhysicalPlan()
-    val result = fun(plan)
-    finalPlanUpdate
-    result
-  }
-
   protected override def stringArgs: Iterator[Any] = Iterator(s"isFinalPlan=$isFinalPlan")
 
   override def generateTreeString(
@@ -521,6 +513,66 @@ case class AdaptiveSparkPlanExec(
     this.inputPlan == obj.asInstanceOf[AdaptiveSparkPlanExec].inputPlan
   }
 
+  /**
+   * We separate stage creation of result and non-result stages because there are several edge cases
+   * of result stage creation:
+   * - existing ResultQueryStage created in previous `withFinalPlanUpdate`.
+   * - the root node is a non-result query stage and we have to create query result stage on top of
+   *   it.
+   * - we create a non-result query stage as root node and the stage is immediately materialized
+   *   due to stage resue, therefore we have to create a result stage right after.
+   *
+   * This method wraps around `createNonResultQueryStages`, the general logic is:
+   * - Early return if ResultQueryStageExec already created before.
+   * - Create non result query stage if possible.
+   * - Try to create result query stage when there is no new non-result query stage created and all
+   *   stages are materialized.
+   */
+  private def createQueryStages(
+      resultHandler: SparkPlan => Any,
+      plan: SparkPlan,
+      firstRun: Boolean): CreateStageResult = {
+    plan match {
+      // 1. ResultQueryStageExec is already created, no need to create non-result stages
+      case resultStage @ ResultQueryStageExec(_, optimizedPlan, _) =>
+        assertStageNotFailed(resultStage)
+        if (firstRun) {
+          // There is already an existing ResultQueryStage created in previous `withFinalPlanUpdate`
+          // e.g, when we do `df.collect` multiple times. Here we create a new result stage to
+          // execute it again, as the handler function can be different.
+          val newResultStage = ResultQueryStageExec(currentStageId, optimizedPlan, resultHandler)
+          currentStageId += 1
+          setLogicalLinkForNewQueryStage(newResultStage, optimizedPlan)
+          CreateStageResult(newPlan = newResultStage,
+            allChildStagesMaterialized = false,
+            newStages = Seq(newResultStage))
+        } else {
+          // We will hit this branch after we've created result query stage in the AQE loop, we
+          // should do nothing.
+          CreateStageResult(newPlan = resultStage,
+            allChildStagesMaterialized = resultStage.isMaterialized,
+            newStages = Seq.empty)
+        }
+      case _ =>
+        // 2. Create non result query stage
+        val result = createNonResultQueryStages(plan)
+        var allNewStages = result.newStages
+        var newPlan = result.newPlan
+        var allChildStagesMaterialized = result.allChildStagesMaterialized
+        // 3. Create result stage
+        if (allNewStages.isEmpty && allChildStagesMaterialized) {
+          val resultStage = newResultQueryStage(resultHandler, newPlan)
+          newPlan = resultStage
+          allChildStagesMaterialized = false
+          allNewStages :+= resultStage
+        }
+        CreateStageResult(
+          newPlan = newPlan,
+          allChildStagesMaterialized = allChildStagesMaterialized,
+          newStages = allNewStages)
+    }
+  }
+
   /**
    * This method is called recursively to traverse the plan tree bottom-up and create a new query
    * stage or try reusing an existing stage if the current node is an [[Exchange]] node and all of
@@ -531,7 +583,7 @@ case class AdaptiveSparkPlanExec(
    * 2) Whether the child query stages (if any) of the current node have all been materialized.
    * 3) A list of the new query stages that have been created.
    */
-  private def createQueryStages(plan: SparkPlan): CreateStageResult = plan match {
+  private def createNonResultQueryStages(plan: SparkPlan): CreateStageResult = plan match {
     case e: Exchange =>
       // First have a quick check in the `stageCache` without having to traverse down the node.
       context.stageCache.get(e.canonicalized) match {
@@ -544,7 +596,7 @@ case class AdaptiveSparkPlanExec(
             newStages = if (isMaterialized) Seq.empty else Seq(stage))
 
         case _ =>
-          val result = createQueryStages(e.child)
+          val result = createNonResultQueryStages(e.child)
           val newPlan = e.withNewChildren(Seq(result.newPlan)).asInstanceOf[Exchange]
           // Create a query stage only when all the child query stages are ready.
           if (result.allChildStagesMaterialized) {
@@ -588,14 +640,28 @@ case class AdaptiveSparkPlanExec(
       if (plan.children.isEmpty) {
         CreateStageResult(newPlan = plan, allChildStagesMaterialized = true, newStages = Seq.empty)
       } else {
-        val results = plan.children.map(createQueryStages)
+        val results = plan.children.map(createNonResultQueryStages)
         CreateStageResult(
           newPlan = plan.withNewChildren(results.map(_.newPlan)),
           allChildStagesMaterialized = results.forall(_.allChildStagesMaterialized),
           newStages = results.flatMap(_.newStages))
       }
   }
 
+  private def newResultQueryStage(
+      resultHandler: SparkPlan => Any,
+      plan: SparkPlan): ResultQueryStageExec = {
+    // Run the final plan when there's no more unfinished stages.
+    val optimizedRootPlan = applyPhysicalRules(
+      optimizeQueryStage(plan, isFinalStage = true),
+      postStageCreationRules(supportsColumnar),
+      Some((planChangeLogger, "AQE Post Stage Creation")))
+    val resultStage = ResultQueryStageExec(currentStageId, optimizedRootPlan, resultHandler)
+    currentStageId += 1
+    setLogicalLinkForNewQueryStage(resultStage, plan)
+    resultStage
+  }
+
   private def newQueryStage(plan: SparkPlan): QueryStageExec = {
     val queryStage = plan match {
       case e: Exchange =>

sql/core/src/main/scala/org/apache/spark/sql/execution/adaptive/AdaptiveSparkPlanHelper.scala

@@ -129,10 +129,12 @@ trait AdaptiveSparkPlanHelper {
   }
 
   /**
-   * Strip the executePlan of AdaptiveSparkPlanExec leaf node.
+   * Strip the top [[AdaptiveSparkPlanExec]] and [[ResultQueryStageExec]] nodes off
+   * the [[SparkPlan]].
    */
   def stripAQEPlan(p: SparkPlan): SparkPlan = p match {
-    case a: AdaptiveSparkPlanExec => a.executedPlan
+    case a: AdaptiveSparkPlanExec => stripAQEPlan(a.executedPlan)
+    case ResultQueryStageExec(_, plan, _) => plan
     case other => other
   }
 }

sql/core/src/main/scala/org/apache/spark/sql/execution/adaptive/QueryStageExec.scala

@@ -19,7 +19,9 @@ package org.apache.spark.sql.execution.adaptive
 
 import java.util.concurrent.atomic.AtomicReference
 
+import scala.concurrent.ExecutionContext
 import scala.concurrent.Future
+import scala.concurrent.Promise
 
 import org.apache.spark.{MapOutputStatistics, SparkException}
 import org.apache.spark.broadcast.Broadcast
@@ -32,7 +34,10 @@ import org.apache.spark.sql.columnar.CachedBatch
 import org.apache.spark.sql.execution._
 import org.apache.spark.sql.execution.columnar.InMemoryTableScanLike
 import org.apache.spark.sql.execution.exchange._
+import org.apache.spark.sql.internal.SQLConf
+import org.apache.spark.sql.internal.StaticSQLConf
 import org.apache.spark.sql.vectorized.ColumnarBatch
+import org.apache.spark.util.ThreadUtils
 
 /**
  * A query stage is an independent subgraph of the query plan. AQE framework will materialize its
@@ -303,3 +308,43 @@ case class TableCacheQueryStageExec(
 
   override def getRuntimeStatistics: Statistics = inMemoryTableScan.runtimeStatistics
 }
+
+case class ResultQueryStageExec(
+    override val id: Int,
+    override val plan: SparkPlan,
+    resultHandler: SparkPlan => Any) extends QueryStageExec {
+
+  override def resetMetrics(): Unit = {
+    plan.resetMetrics()
+  }
+
+  override protected def doMaterialize(): Future[Any] = {
+    val javaFuture = SQLExecution.withThreadLocalCaptured(
+      session,
+      ResultQueryStageExec.executionContext) {
+      resultHandler(plan)
+    }
+    val scalaPromise: Promise[Any] = Promise()
+    javaFuture.whenComplete { (result: Any, exception: Throwable) =>
+      if (exception != null) {
+        scalaPromise.failure(exception match {
+          case completionException: java.util.concurrent.CompletionException =>
+            completionException.getCause
+          case ex => ex
+        })
+      } else {
+        scalaPromise.success(result)
+      }
+    }
+    scalaPromise.future
+  }
+
+  // Result stage could be any SparkPlan, so we don't have a specific runtime statistics for it.
+  override def getRuntimeStatistics: Statistics = Statistics.DUMMY
+}
+
+object ResultQueryStageExec {
+  private[execution] val executionContext = ExecutionContext.fromExecutorService(
+    ThreadUtils.newDaemonCachedThreadPool("ResultQueryStageExecution",
+      SQLConf.get.getConf(StaticSQLConf.RESULT_QUERY_STAGE_MAX_THREAD_THRESHOLD)))
+}

sql/core/src/main/scala/org/apache/spark/sql/execution/ui/SparkPlanGraph.scala

@@ -106,7 +106,7 @@ object SparkPlanGraph {
           buildSparkPlanGraphNode(
             planInfo.children.head, nodeIdGenerator, nodes, edges, parent, null, exchanges)
         }
-      case "TableCacheQueryStage" =>
+      case "TableCacheQueryStage" | "ResultQueryStage" =>
         buildSparkPlanGraphNode(
           planInfo.children.head, nodeIdGenerator, nodes, edges, parent, null, exchanges)
       case "Subquery" if subgraph != null =>

sql/core/src/test/scala/org/apache/spark/sql/CachedTableSuite.scala

@@ -1659,7 +1659,9 @@ class CachedTableSuite extends QueryTest with SQLTestUtils
           _.nodeName.contains("TableCacheQueryStage"))
         val aqeNode = findNodeInSparkPlanInfo(inMemoryScanNode.get,
           _.nodeName.contains("AdaptiveSparkPlan"))
-        aqeNode.get.children.head.nodeName == "AQEShuffleRead"
+        val aqePlanRoot = findNodeInSparkPlanInfo(inMemoryScanNode.get,
+          _.nodeName.contains("ResultQueryStage"))
+        aqePlanRoot.get.children.head.nodeName == "AQEShuffleRead"
       }
 
       withTempView("t0", "t1", "t2") {