[SPARK-20540][CORE] Fix unstable executor requests.

There are two problems fixed in this commit. First, the
ExecutorAllocationManager sets a timeout to avoid requesting executors
too often. However, the timeout is always updated based on its value and
a timeout, not the current time. If the call is delayed by locking for
more than the ongoing scheduler timeout, the manager will request more
executors on every run. This seems to be the main cause of SPARK-20540.

The second problem is that the total number of requested executors is
not tracked by the CoarseGrainedSchedulerBackend. Instead, it calculates
the value based on the current status of 3 variables: the number of
known executors, the number of executors that have been killed, and the
number of pending executors. But, the number of pending executors is
never less than 0, even though there may be more known than requested.
When executors are killed and not replaced, this can cause the request
sent to YARN to be incorrect because there were too many executors due
to the scheduler's state being slightly out of date. This is fixed by tracking
the currently requested size explicitly.

## How was this patch tested?

Existing tests.

Author: Ryan Blue <[email protected]>

Closes #17813 from rdblue/SPARK-20540-fix-dynamic-allocation.

Author: rdblue

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

@@ -331,7 +331,7 @@ private[spark] class ExecutorAllocationManager(
       val delta = addExecutors(maxNeeded)
       logDebug(s"Starting timer to add more executors (to " +
         s"expire in $sustainedSchedulerBacklogTimeoutS seconds)")
-      addTime += sustainedSchedulerBacklogTimeoutS * 1000
+      addTime = now + (sustainedSchedulerBacklogTimeoutS * 1000)
       delta
     } else {
       0

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

@@ -69,6 +69,10 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
   // `CoarseGrainedSchedulerBackend.this`.
   private val executorDataMap = new HashMap[String, ExecutorData]
 
+  // Number of executors requested by the cluster manager, [[ExecutorAllocationManager]]
+  @GuardedBy("CoarseGrainedSchedulerBackend.this")
+  private var requestedTotalExecutors = 0
+
   // Number of executors requested from the cluster manager that have not registered yet
   @GuardedBy("CoarseGrainedSchedulerBackend.this")
   private var numPendingExecutors = 0
@@ -413,6 +417,7 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
    * */
   protected def reset(): Unit = {
     val executors = synchronized {
+      requestedTotalExecutors = 0
       numPendingExecutors = 0
       executorsPendingToRemove.clear()
       Set() ++ executorDataMap.keys
@@ -487,12 +492,21 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
     logInfo(s"Requesting $numAdditionalExecutors additional executor(s) from the cluster manager")
 
     val response = synchronized {
+      requestedTotalExecutors += numAdditionalExecutors
       numPendingExecutors += numAdditionalExecutors
       logDebug(s"Number of pending executors is now $numPendingExecutors")
+      if (requestedTotalExecutors !=
+          (numExistingExecutors + numPendingExecutors - executorsPendingToRemove.size)) {
+        logDebug(
+          s"""requestExecutors($numAdditionalExecutors): Executor request doesn't match:
+             |requestedTotalExecutors  = $requestedTotalExecutors
+             |numExistingExecutors     = $numExistingExecutors
+             |numPendingExecutors      = $numPendingExecutors
+             |executorsPendingToRemove = ${executorsPendingToRemove.size}""".stripMargin)
+      }
 
       // Account for executors pending to be added or removed
-      doRequestTotalExecutors(
-        numExistingExecutors + numPendingExecutors - executorsPendingToRemove.size)
+      doRequestTotalExecutors(requestedTotalExecutors)
     }
 
     defaultAskTimeout.awaitResult(response)
@@ -524,6 +538,7 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
     }
 
     val response = synchronized {
+      this.requestedTotalExecutors = numExecutors
       this.localityAwareTasks = localityAwareTasks
       this.hostToLocalTaskCount = hostToLocalTaskCount
 
@@ -589,8 +604,17 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
       // take into account executors that are pending to be added or removed.
       val adjustTotalExecutors =
         if (!replace) {
-          doRequestTotalExecutors(
-            numExistingExecutors + numPendingExecutors - executorsPendingToRemove.size)
+          requestedTotalExecutors = math.max(requestedTotalExecutors - executorsToKill.size, 0)
+          if (requestedTotalExecutors !=
+              (numExistingExecutors + numPendingExecutors - executorsPendingToRemove.size)) {
+            logDebug(
+              s"""killExecutors($executorIds, $replace, $force): Executor counts do not match:
+                 |requestedTotalExecutors  = $requestedTotalExecutors
+                 |numExistingExecutors     = $numExistingExecutors
+                 |numPendingExecutors      = $numPendingExecutors
+                 |executorsPendingToRemove = ${executorsPendingToRemove.size}""".stripMargin)
+          }
+          doRequestTotalExecutors(requestedTotalExecutors)
         } else {
           numPendingExecutors += knownExecutors.size
           Future.successful(true)

core/src/test/scala/org/apache/spark/deploy/StandaloneDynamicAllocationSuite.scala

@@ -356,12 +356,13 @@ class StandaloneDynamicAllocationSuite
   test("kill the same executor twice (SPARK-9795)") {
     sc = new SparkContext(appConf)
     val appId = sc.applicationId
+    sc.requestExecutors(2)
     eventually(timeout(10.seconds), interval(10.millis)) {
       val apps = getApplications()
       assert(apps.size === 1)
       assert(apps.head.id === appId)
       assert(apps.head.executors.size === 2)
-      assert(apps.head.getExecutorLimit === Int.MaxValue)
+      assert(apps.head.getExecutorLimit === 2)
     }
     // sync executors between the Master and the driver, needed because
     // the driver refuses to kill executors it does not know about
@@ -380,12 +381,13 @@ class StandaloneDynamicAllocationSuite
   test("the pending replacement executors should not be lost (SPARK-10515)") {
     sc = new SparkContext(appConf)
     val appId = sc.applicationId
+    sc.requestExecutors(2)
     eventually(timeout(10.seconds), interval(10.millis)) {
       val apps = getApplications()
       assert(apps.size === 1)
       assert(apps.head.id === appId)
       assert(apps.head.executors.size === 2)
-      assert(apps.head.getExecutorLimit === Int.MaxValue)
+      assert(apps.head.getExecutorLimit === 2)
     }
     // sync executors between the Master and the driver, needed because
     // the driver refuses to kill executors it does not know about