Connect decommissioning to dynamic scaling

Because the mock always says there is an RDD we may replicate more than once, and now that there are independent threads

Make Spark's dynamic allocation use decommissioning

Track the decommissioning executors in the core dynamic scheduler so we don't scale down too low, update the streaming ExecutorAllocationManager to also delegate to decommission

Fix up executor add for resource profile

Fix our exiting and cleanup thread for better debugging next time. Cleanup the locks we use in decommissioning and clarify some more bits.

Verify executors decommissioned, then killed by external external cluster manager are re-launched

Verify some additional calls are not occuring in the executor allocation manager suite.

Dont' close the watcher until the end of the test

Use decommissionExecutors and set adjustTargetNumExecutors to false so that we can match the pattern for killExecutor/killExecutors

bump numparts up to 6

Revert "bump numparts up to 6"

This reverts commit daf96dd.

Small coment & visibility cleanup

Author: holdenk

core/src/main/scala/org/apache/spark/ExecutorAllocationClient.scala

@@ -17,6 +17,7 @@
 
 package org.apache.spark
 
+import org.apache.spark.scheduler.ExecutorDecommissionInfo
 /**
  * A client that communicates with the cluster manager to request or kill executors.
  * This is currently supported only in YARN mode.
@@ -81,6 +82,43 @@ private[spark] trait ExecutorAllocationClient {
     countFailures: Boolean,
     force: Boolean = false): Seq[String]
 
+  /**
+   * Request that the cluster manager decommission the specified executors.
+   * Default implementation delegates to kill, scheduler must override
+   * if it supports graceful decommissioning.
+   *
+   * @param executors identifiers of executors & decom info.
+   * @param adjustTargetNumExecutors whether the target number of executors will be adjusted down
+   *                                 after these executors have been decommissioned.
+   * @return the ids of the executors acknowledged by the cluster manager to be removed.
+   */
+  def decommissionExecutors(
+    executors: Seq[(String, ExecutorDecommissionInfo)],
+    adjustTargetNumExecutors: Boolean): Seq[String] = {
+    killExecutors(executors.map(_._1),
+      adjustTargetNumExecutors,
+      countFailures = false)
+  }
+
+
+  /**
+   * Request that the cluster manager decommission the specified executor.
+   * Default implementation delegates to decommissionExecutors, scheduler can override
+   * if it supports graceful decommissioning.
+   *
+   * @param executorId identifiers of executor to decommission
+   * @param decommissionInfo information about the decommission (reason, host loss)
+   * @return whether the request is acknowledged by the cluster manager.
+   */
+  def decommissionExecutor(executorId: String,
+      decommissionInfo: ExecutorDecommissionInfo): Boolean = {
+    val decommissionedExecutors = decommissionExecutors(
+      Seq((executorId, decommissionInfo)),
+      adjustTargetNumExecutors = true)
+    decommissionedExecutors.nonEmpty && decommissionedExecutors(0).equals(executorId)
+  }
+
+
   /**
    * Request that the cluster manager kill every executor on the specified host.
    *

core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala

@@ -28,6 +28,7 @@ import com.codahale.metrics.{Gauge, MetricRegistry}
 import org.apache.spark.internal.{config, Logging}
 import org.apache.spark.internal.config._
 import org.apache.spark.internal.config.Tests.TEST_SCHEDULE_INTERVAL
+import org.apache.spark.internal.config.Worker.WORKER_DECOMMISSION_ENABLED
 import org.apache.spark.metrics.source.Source
 import org.apache.spark.resource.ResourceProfile.UNKNOWN_RESOURCE_PROFILE_ID
 import org.apache.spark.resource.ResourceProfileManager
@@ -204,7 +205,12 @@ private[spark] class ExecutorAllocationManager(
         s"s${DYN_ALLOCATION_SUSTAINED_SCHEDULER_BACKLOG_TIMEOUT.key} must be > 0!")
     }
     if (!conf.get(config.SHUFFLE_SERVICE_ENABLED)) {
-      if (conf.get(config.DYN_ALLOCATION_SHUFFLE_TRACKING_ENABLED)) {
+      // If dynamic allocation shuffle tracking or worker decommissioning along with
+      // storage shuffle decommissioning is enabled we have *experimental* support for
+      // decommissioning without a shuffle service.
+      if (conf.get(config.DYN_ALLOCATION_SHUFFLE_TRACKING_ENABLED) ||
+          (conf.get(WORKER_DECOMMISSION_ENABLED) &&
+            conf.get(config.STORAGE_DECOMMISSION_SHUFFLE_BLOCKS_ENABLED))) {
         logWarning("Dynamic allocation without a shuffle service is an experimental feature.")
       } else if (!testing) {
         throw new SparkException("Dynamic allocation of executors requires the external " +
@@ -539,7 +545,9 @@ private[spark] class ExecutorAllocationManager(
         // get the running total as we remove or initialize it to the count - pendingRemoval
         val newExecutorTotal = numExecutorsTotalPerRpId.getOrElseUpdate(rpId,
           (executorMonitor.executorCountWithResourceProfile(rpId) -
-            executorMonitor.pendingRemovalCountPerResourceProfileId(rpId)))
+            executorMonitor.pendingRemovalCountPerResourceProfileId(rpId) -
+            executorMonitor.pendingDecommissioningPerResourceProfileId(rpId)
+          ))
         if (newExecutorTotal - 1 < minNumExecutors) {
           logDebug(s"Not removing idle executor $executorIdToBeRemoved because there " +
             s"are only $newExecutorTotal executor(s) left (minimum number of executor limit " +
@@ -565,8 +573,14 @@ private[spark] class ExecutorAllocationManager(
     } else {
       // We don't want to change our target number of executors, because we already did that
       // when the task backlog decreased.
-      client.killExecutors(executorIdsToBeRemoved.toSeq, adjustTargetNumExecutors = false,
-        countFailures = false, force = false)
+      if (conf.get(WORKER_DECOMMISSION_ENABLED)) {
+        val executorIdsWithoutHostLoss = executorIdsToBeRemoved.toSeq.map(
+          id => (id, ExecutorDecommissionInfo("spark scale down", false)))
+        client.decommissionExecutors(executorIdsWithoutHostLoss, adjustTargetNumExecutors = false)
+      } else {
+        client.killExecutors(executorIdsToBeRemoved.toSeq, adjustTargetNumExecutors = false,
+          countFailures = false, force = false)
+      }
     }
 
     // [SPARK-21834] killExecutors api reduces the target number of executors.
@@ -578,7 +592,11 @@ private[spark] class ExecutorAllocationManager(
 
     // reset the newExecutorTotal to the existing number of executors
     if (testing || executorsRemoved.nonEmpty) {
-      executorMonitor.executorsKilled(executorsRemoved.toSeq)
+      if (conf.get(WORKER_DECOMMISSION_ENABLED)) {
+        executorMonitor.executorsDecommissioned(executorsRemoved)
+      } else {
+        executorMonitor.executorsKilled(executorsRemoved.toSeq)
+      }
       logInfo(s"Executors ${executorsRemoved.mkString(",")} removed due to idle timeout.")
       executorsRemoved.toSeq
     } else {

core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala

@@ -193,7 +193,8 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
 
       case DecommissionExecutor(executorId, decommissionInfo) =>
         logError(s"Received decommission executor message ${executorId}: $decommissionInfo")
-        decommissionExecutor(executorId, decommissionInfo)
+        decommissionExecutors(Seq((executorId, decommissionInfo)),
+          adjustTargetNumExecutors = false)
 
       case RemoveWorker(workerId, host, message) =>
         removeWorker(workerId, host, message)
@@ -274,7 +275,8 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
 
       case DecommissionExecutor(executorId, decommissionInfo) =>
         logError(s"Received decommission executor message ${executorId}: ${decommissionInfo}.")
-        decommissionExecutor(executorId, decommissionInfo)
+        decommissionExecutors(Seq((executorId, decommissionInfo)),
+          adjustTargetNumExecutors = false)
         context.reply(true)
 
       case RetrieveSparkAppConfig(resourceProfileId) =>
@@ -503,6 +505,100 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
 
   protected def minRegisteredRatio: Double = _minRegisteredRatio
 
+  /**
+   * Request that the cluster manager decommission the specified executors.
+   *
+   * @param executors Identifiers of executors & decommission info.
+   * @param adjustTargetNumExecutors whether the target number of executors will be adjusted down
+   *                                 after these executors have been decommissioned.
+   * @return the ids of the executors acknowledged by the cluster manager to be removed.
+   */
+  override def decommissionExecutors(
+      executors: Seq[(String, ExecutorDecommissionInfo)],
+      adjustTargetNumExecutors: Boolean): Seq[String] = {
+
+    val executorsToDecommission = executors.filter{case (executorId, _) =>
+      CoarseGrainedSchedulerBackend.this.synchronized {
+        // Only bother decommissioning executors which are alive.
+        if (isExecutorActive(executorId)) {
+          executorsPendingDecommission += executorId
+          true
+        } else {
+          false
+        }
+      }
+    }
+
+    // If we don't want to replace the executors we are decommissioning
+    if (adjustTargetNumExecutors) {
+      executorsToDecommission.foreach { case (exec, _) =>
+        val rpId = executorDataMap(exec).resourceProfileId
+        val rp = scheduler.sc.resourceProfileManager.resourceProfileFromId(rpId)
+        if (requestedTotalExecutorsPerResourceProfile.isEmpty) {
+          // Assume that we are killing an executor that was started by default and
+          // not through the request api
+          requestedTotalExecutorsPerResourceProfile(rp) = 0
+        } else {
+          val requestedTotalForRp = requestedTotalExecutorsPerResourceProfile(rp)
+          requestedTotalExecutorsPerResourceProfile(rp) = math.max(requestedTotalForRp - 1, 0)
+        }
+      }
+      doRequestTotalExecutors(requestedTotalExecutorsPerResourceProfile.toMap)
+    }
+
+    val decommissioned = executorsToDecommission.filter{case (executorId, decomInfo) =>
+      doDecommission(executorId, decomInfo)
+    }.map(_._1)
+    decommissioned
+  }
+
+
+  private def doDecommission(executorId: String,
+      decomInfo: ExecutorDecommissionInfo): Boolean = {
+
+    logInfo(s"Starting decommissioning executor $executorId.")
+    try {
+      scheduler.executorDecommission(executorId, decomInfo)
+      if (driverEndpoint != null) {
+        logInfo("Propagating executor decommission to driver.")
+        driverEndpoint.send(DecommissionExecutor(executorId, decomInfo))
+      }
+    } catch {
+      case e: Exception =>
+        logError(s"Unexpected error during decommissioning ${e.toString}", e)
+        return false
+    }
+    // Send decommission message to the executor (it could have originated on the executor
+    // but not necessarily.
+    CoarseGrainedSchedulerBackend.this.synchronized {
+      executorDataMap.get(executorId) match {
+        case Some(executorInfo) =>
+          executorInfo.executorEndpoint.send(DecommissionSelf)
+        case None =>
+          // Ignoring the executor since it is not registered.
+          logWarning(s"Attempted to decommission unknown executor $executorId.")
+          return false
+      }
+    }
+    logInfo(s"Finished decommissioning executor $executorId.")
+
+    if (conf.get(STORAGE_DECOMMISSION_ENABLED)) {
+      try {
+        logInfo("Starting decommissioning block manager corresponding to " +
+          s"executor $executorId.")
+        scheduler.sc.env.blockManager.master.decommissionBlockManagers(Seq(executorId))
+      } catch {
+        case e: Exception =>
+          logError("Unexpected error during block manager " +
+            s"decommissioning for executor $executorId: ${e.toString}", e)
+          return false
+      }
+      logInfo(s"Acknowledged decommissioning block manager corresponding to $executorId.")
+    }
+    true
+  }
+
+
   override def start(): Unit = {
     if (UserGroupInformation.isSecurityEnabled()) {
       delegationTokenManager = createTokenManager()
@@ -598,17 +694,6 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
     driverEndpoint.send(RemoveWorker(workerId, host, message))
   }
 
-  /**
-   * Called by subclasses when notified of a decommissioning executor.
-   */
-  private[spark] def decommissionExecutor(
-      executorId: String, decommissionInfo: ExecutorDecommissionInfo): Unit = {
-    if (driverEndpoint != null) {
-      logInfo("Propagating executor decommission to driver.")
-      driverEndpoint.send(DecommissionExecutor(executorId, decommissionInfo))
-    }
-  }
-
   def sufficientResourcesRegistered(): Boolean = true
 
   override def isReady(): Boolean = {

core/src/main/scala/org/apache/spark/scheduler/cluster/StandaloneSchedulerBackend.scala

@@ -176,7 +176,8 @@ private[spark] class StandaloneSchedulerBackend(
 
   override def executorDecommissioned(fullId: String, decommissionInfo: ExecutorDecommissionInfo) {
     logInfo("Asked to decommission executor")
-    decommissionExecutor(fullId.split("/")(1), decommissionInfo)
+    val execId = fullId.split("/")(1)
+    decommissionExecutors(Seq((execId, decommissionInfo)), adjustTargetNumExecutors = false)
     logInfo("Executor %s decommissioned: %s".format(fullId, decommissionInfo))
   }
 

core/src/main/scala/org/apache/spark/scheduler/dynalloc/ExecutorMonitor.scala

@@ -28,7 +28,7 @@ import org.apache.spark.internal.Logging
 import org.apache.spark.internal.config._
 import org.apache.spark.resource.ResourceProfile.UNKNOWN_RESOURCE_PROFILE_ID
 import org.apache.spark.scheduler._
-import org.apache.spark.storage.RDDBlockId
+import org.apache.spark.storage.{RDDBlockId, ShuffleDataBlockId}
 import org.apache.spark.util.Clock
 
 /**
@@ -114,7 +114,8 @@ private[spark] class ExecutorMonitor(
 
       var newNextTimeout = Long.MaxValue
       timedOutExecs = executors.asScala
-        .filter { case (_, exec) => !exec.pendingRemoval && !exec.hasActiveShuffle }
+        .filter { case (_, exec) =>
+          !exec.pendingRemoval && !exec.hasActiveShuffle && !exec.pendingDecommissioning}
         .filter { case (_, exec) =>
           val deadline = exec.timeoutAt
           if (deadline > now) {
@@ -135,6 +136,7 @@ private[spark] class ExecutorMonitor(
 
   /**
    * Mark the given executors as pending to be removed. Should only be called in the EAM thread.
+   * This covers both kills and decommissions.
    */
   def executorsKilled(ids: Seq[String]): Unit = {
     ids.foreach { id =>
@@ -149,6 +151,19 @@ private[spark] class ExecutorMonitor(
     nextTimeout.set(Long.MinValue)
   }
 
+  private[spark] def executorsDecommissioned(ids: Seq[String]): Unit = {
+    ids.foreach { id =>
+      val tracker = executors.get(id)
+      if (tracker != null) {
+        tracker.pendingDecommissioning = true
+      }
+    }
+
+    // Recompute timed out executors in the next EAM callback, since this call invalidates
+    // the current list.
+    nextTimeout.set(Long.MinValue)
+  }
+
   def executorCount: Int = executors.size()
 
   def executorCountWithResourceProfile(id: Int): Int = {
@@ -171,6 +186,16 @@ private[spark] class ExecutorMonitor(
     executors.asScala.filter { case (k, v) => v.resourceProfileId == id && v.pendingRemoval }.size
   }
 
+  def pendingDecommissioningCount: Int = executors.asScala.count { case (_, exec) =>
+    exec.pendingDecommissioning
+  }
+
+  def pendingDecommissioningPerResourceProfileId(id: Int): Int = {
+    executors.asScala.filter { case (k, v) =>
+      v.resourceProfileId == id && v.pendingDecommissioning
+    }.size
+  }
+
   override def onJobStart(event: SparkListenerJobStart): Unit = {
     if (!shuffleTrackingEnabled) {
       return
@@ -298,6 +323,7 @@ private[spark] class ExecutorMonitor(
       //
       // This means that an executor may be marked as having shuffle data, and thus prevented
       // from being removed, even though the data may not be used.
+      // TODO: Only track used files (SPARK-31974)
       if (shuffleTrackingEnabled && event.reason == Success) {
         stageToShuffleID.get(event.stageId).foreach { shuffleId =>
           exec.addShuffle(shuffleId)
@@ -326,18 +352,33 @@ private[spark] class ExecutorMonitor(
     val removed = executors.remove(event.executorId)
     if (removed != null) {
       decrementExecResourceProfileCount(removed.resourceProfileId)
-      if (!removed.pendingRemoval) {
+      if (!removed.pendingRemoval || !removed.pendingDecommissioning) {
         nextTimeout.set(Long.MinValue)
       }
     }
   }
 
   override def onBlockUpdated(event: SparkListenerBlockUpdated): Unit = {
-    if (!event.blockUpdatedInfo.blockId.isInstanceOf[RDDBlockId]) {
-      return
-    }
     val exec = ensureExecutorIsTracked(event.blockUpdatedInfo.blockManagerId.executorId,
       UNKNOWN_RESOURCE_PROFILE_ID)
+
+    // Check if it is a shuffle file, or RDD to pick the correct codepath for update
+    if (event.blockUpdatedInfo.blockId.isInstanceOf[ShuffleDataBlockId] && shuffleTrackingEnabled) {
+      /**
+       * The executor monitor keeps track of locations of cache and shuffle blocks and this can be
+       * used to decide which executor(s) Spark should shutdown first. Since we move shuffle blocks
+       * around now this wires it up so that it keeps track of it. We only do this for data blocks
+       * as index and other blocks blocks do not necessarily mean the entire block has been
+       * committed.
+       */
+      event.blockUpdatedInfo.blockId match {
+        case ShuffleDataBlockId(shuffleId, _, _) => exec.addShuffle(shuffleId)
+        case _ => // For now we only update on data blocks
+      }
+      return
+    } else if (!event.blockUpdatedInfo.blockId.isInstanceOf[RDDBlockId]) {
+      return
+    }
     val storageLevel = event.blockUpdatedInfo.storageLevel
     val blockId = event.blockUpdatedInfo.blockId.asInstanceOf[RDDBlockId]
 
@@ -410,10 +451,15 @@ private[spark] class ExecutorMonitor(
   }
 
   // Visible for testing
-  def executorsPendingToRemove(): Set[String] = {
+  private[spark] def executorsPendingToRemove(): Set[String] = {
     executors.asScala.filter { case (_, exec) => exec.pendingRemoval }.keys.toSet
   }
 
+  // Visible for testing
+  private[spark] def executorsDecommissioning(): Set[String] = {
+    executors.asScala.filter { case (_, exec) => exec.pendingDecommissioning }.keys.toSet
+  }
+
   /**
    * This method should be used when updating executor state. It guards against a race condition in
    * which the `SparkListenerTaskStart` event is posted before the `SparkListenerBlockManagerAdded`
@@ -466,6 +512,7 @@ private[spark] class ExecutorMonitor(
     @volatile var timedOut: Boolean = false
 
     var pendingRemoval: Boolean = false
+    var pendingDecommissioning: Boolean = false
     var hasActiveShuffle: Boolean = false
 
     private var idleStart: Long = -1