Reduce job failures during decommissioning

This PR reduces the prospect of a job loss during decommissioning. It
fixes two holes in the current decommissioning framework:

- (a) Loss of decommissioned executors is not treated as a job failure:
We know that the decom'd executor would be dying soon, so its death is
clearly not caused by the application.

- (b) Shuffle files on the decommissioned host are cleared when the
first fetch failure is detected from a decommissioned host: This is a
bit tricky in terms of when to clear the shuffle state ? Ideally you
want to clear it the millisecond before the shuffle service on the node
dies (or the executor dies when there is no external shuffle service) --
too soon and it could lead to some wasteage and too late would lead to
fetch failures.

(These two things are a bit intertwined so it is easier to do them in a
single commit)

The approach here is to do this clearing when the very first fetch
failure is observed on the decom'd block manager, without waiting for
other blocks to also signal a failure.

Added a new unit test `DecommissionWorkerSuite` to test the new
behavior.

Author: dagrawal3409

core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala

@@ -1821,10 +1821,19 @@ private[spark] class DAGScheduler(
 
           // TODO: mark the executor as failed only if there were lots of fetch failures on it
           if (bmAddress != null) {
-            val hostToUnregisterOutputs = if (env.blockManager.externalShuffleServiceEnabled &&
-              unRegisterOutputOnHostOnFetchFailure) {
-              // We had a fetch failure with the external shuffle service, so we
-              // assume all shuffle data on the node is bad.
+            val externalShuffleServiceEnabled = env.blockManager.externalShuffleServiceEnabled
+            val isHostDecommissioned = taskScheduler
+              .getExecutorDecommissionInfo(bmAddress.executorId)
+              .exists(_.isHostDecommissioned)
+
+            // Shuffle output of all executors on host `bmAddress.host` may be lost if:
+            // - External shuffle service is enabled, so we assume that all shuffle data on node is
+            //   bad.
+            // - Host is decommissioned, thus all executors on that host will die.
+            val shuffleOutputOfEntireHostLost = externalShuffleServiceEnabled ||
+              isHostDecommissioned
+            val hostToUnregisterOutputs = if (shuffleOutputOfEntireHostLost
+              && unRegisterOutputOnHostOnFetchFailure) {
               Some(bmAddress.host)
             } else {
               // Unregister shuffle data just for one executor (we don't have any
@@ -2339,7 +2348,7 @@ private[scheduler] class DAGSchedulerEventProcessLoop(dagScheduler: DAGScheduler
 
     case ExecutorLost(execId, reason) =>
       val workerLost = reason match {
-        case ExecutorProcessLost(_, true) => true
+        case ExecutorProcessLost(_, true, _) => true
         case _ => false
       }
       dagScheduler.handleExecutorLost(execId, workerLost)

core/src/main/scala/org/apache/spark/scheduler/ExecutorLossReason.scala

@@ -54,9 +54,14 @@ private [spark] object LossReasonPending extends ExecutorLossReason("Pending los
 /**
  * @param _message human readable loss reason
  * @param workerLost whether the worker is confirmed lost too (i.e. including shuffle service)
+ * @param causedByApp whether the loss of the executor is the fault of the running app.
+ *                    (assumed true by default unless known explicitly otherwise)
  */
 private[spark]
-case class ExecutorProcessLost(_message: String = "Worker lost", workerLost: Boolean = false)
+case class ExecutorProcessLost(
+    _message: String = "Executor Process Lost",
+    workerLost: Boolean = false,
+    causedByApp: Boolean = true)
   extends ExecutorLossReason(_message)
 
 /**

core/src/main/scala/org/apache/spark/scheduler/TaskScheduler.scala

@@ -103,6 +103,11 @@ private[spark] trait TaskScheduler {
    */
   def executorDecommission(executorId: String, decommissionInfo: ExecutorDecommissionInfo): Unit
 
+  /**
+   * If an executor is decommissioned, return its corresponding decommission info
+   */
+  def getExecutorDecommissionInfo(executorId: String): Option[ExecutorDecommissionInfo]
+
   /**
    * Process a lost executor
    */

core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala

@@ -136,6 +136,8 @@ private[spark] class TaskSchedulerImpl(
   // IDs of the tasks running on each executor
   private val executorIdToRunningTaskIds = new HashMap[String, HashSet[Long]]
 
+  private val executorsPendingDecommission = new HashMap[String, ExecutorDecommissionInfo]
+
   def runningTasksByExecutors: Map[String, Int] = synchronized {
     executorIdToRunningTaskIds.toMap.mapValues(_.size).toMap
   }
@@ -939,12 +941,43 @@ private[spark] class TaskSchedulerImpl(
 
   override def executorDecommission(
       executorId: String, decommissionInfo: ExecutorDecommissionInfo): Unit = {
+    synchronized {
+      // Don't bother noting decommissioning for executors that we don't know about
+      if (executorIdToHost.contains(executorId)) {
+        // The scheduler can get multiple decommission updates from multiple sources,
+        // and some of those can have isHostDecommissioned false. We merge them such that
+        // if we heard isHostDecommissioned ever true, then we keep that one since it is
+        // most likely coming from the cluster manager and thus authoritative
+        val oldDecomInfo = executorsPendingDecommission.get(executorId)
+        if (oldDecomInfo.isEmpty || !oldDecomInfo.get.isHostDecommissioned) {
+          executorsPendingDecommission(executorId) = decommissionInfo
+        }
+      }
+    }
     rootPool.executorDecommission(executorId)
     backend.reviveOffers()
   }
 
-  override def executorLost(executorId: String, reason: ExecutorLossReason): Unit = {
+  override def getExecutorDecommissionInfo(executorId: String)
+    : Option[ExecutorDecommissionInfo] = synchronized {
+      executorsPendingDecommission.get(executorId)
+  }
+
+  override def executorLost(executorId: String, givenReason: ExecutorLossReason): Unit = {
     var failedExecutor: Option[String] = None
+    val reason = givenReason match {
+      // Handle executor process loss due to decommissioning
+      case ExecutorProcessLost(message, origWorkerLost, origCausedByApp) =>
+        val executorDecommissionInfo = getExecutorDecommissionInfo(executorId)
+        ExecutorProcessLost(
+          message,
+          // Also mark the worker lost if we know that the host was decommissioned
+          origWorkerLost || executorDecommissionInfo.exists(_.isHostDecommissioned),
+          // Executor loss is certainly not caused by app if we knew that this executor is being
+          // decommissioned
+          causedByApp = executorDecommissionInfo.isEmpty && origCausedByApp)
+      case e => e
+    }
 
     synchronized {
       if (executorIdToRunningTaskIds.contains(executorId)) {
@@ -1033,6 +1066,8 @@ private[spark] class TaskSchedulerImpl(
       }
     }
 
+    executorsPendingDecommission -= executorId
+
     if (reason != LossReasonPending) {
       executorIdToHost -= executorId
       rootPool.executorLost(executorId, host, reason)

core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala

@@ -985,6 +985,7 @@ private[spark] class TaskSetManager(
       val exitCausedByApp: Boolean = reason match {
         case exited: ExecutorExited => exited.exitCausedByApp
         case ExecutorKilled => false
+        case ExecutorProcessLost(_, _, false) => false
         case _ => true
       }
       handleFailedTask(tid, TaskState.FAILED, ExecutorLostFailure(info.executorId, exitCausedByApp,