DCOS-57560: Suppress/revive support for Mesos (#67)

DCOS-57560: Suppress/revive support for Mesos

- Suppresses Mesos offers when the executor cap is reached (dynamic allocation enabled)
- Reviving Mesos offers when the executor cap is increased (dynamic allocation enabled)
- Suppresses Mesos offers when max core number is utilized.
- Doesn't revive if a task finished successfully
- Periodically revives Mesos offers if needed  using "spark.mesos.scheduler.revive.interval"

Author: rpalaznik

resource-managers/mesos/src/main/scala/org/apache/spark/deploy/mesos/config.scala

@@ -140,4 +140,11 @@ package object config {
         "when launching drivers. Default is to accept all offers with sufficient resources.")
       .stringConf
       .createWithDefault("")
+
+  private[spark] val REVIVE_OFFERS_INTERVAL =
+    ConfigBuilder("spark.mesos.scheduler.revive.interval")
+      .doc("Amount of milliseconds between periodic revive calls to Mesos, when the job" +
+        "driver is not suppressing resource offers.")
+      .timeConf(TimeUnit.MILLISECONDS)
+      .createWithDefaultString("10s")
 }

resource-managers/mesos/src/main/scala/org/apache/spark/scheduler/cluster/mesos/MesosClusterScheduler.scala

@@ -20,6 +20,8 @@ package org.apache.spark.scheduler.cluster.mesos
 import java.io.File
 import java.util.{Collections, Date, List => JList}
 
+import org.apache.commons.lang3.time.DateUtils
+
 import org.apache.mesos.{Protos, Scheduler, SchedulerDriver}
 import org.apache.mesos.Protos.{TaskState => MesosTaskState, _}
 import org.apache.mesos.Protos.Environment.Variable
@@ -779,19 +781,20 @@ private[spark] class MesosClusterScheduler(
   }
 
   private def getNewRetryState(
-    retryState: Option[MesosClusterRetryState], status: TaskStatus): MesosClusterRetryState = {
-
-    val (retries, waitTimeSec) = retryState
-      .map { rs => (rs.retries + 1, rs.waitTime) }
-      .getOrElse{ (1, 1) }
-
-    // if a node is draining, the driver should be relaunched without backoff
-    if (isNodeDraining(status)) {
-      new MesosClusterRetryState(status, retries, new Date(), waitTimeSec)
-    } else {
-      val newWaitTime = waitTimeSec * 2
-      val nextRetry = new Date(new Date().getTime + newWaitTime * 1000L)
-      new MesosClusterRetryState(status, retries, nextRetry, newWaitTime)
+      retryState: Option[MesosClusterRetryState], status: TaskStatus): MesosClusterRetryState = {
+    val now = new Date()
+    retryState.map { rs =>
+      val newRetries = rs.retries + 1
+      // if a node is draining, the driver should be relaunched without backoff
+      if (isNodeDraining(status)) {
+        new MesosClusterRetryState(status, newRetries, now, rs.waitTime)
+      } else {
+        new MesosClusterRetryState(
+          status, newRetries, DateUtils.addSeconds(now, rs.waitTime), rs.waitTime * 2)
+      }
+    }.getOrElse {
+      // this is the first retry which should happen without backoff
+      new MesosClusterRetryState(status, 1, now, 1)
     }
   }
 

resource-managers/mesos/src/main/scala/org/apache/spark/scheduler/cluster/mesos/MesosCoarseGrainedSchedulerBackend.scala

@@ -19,6 +19,7 @@ package org.apache.spark.scheduler.cluster.mesos
 
 import java.io.File
 import java.util.{Collections, List => JList, UUID}
+import java.util.concurrent.TimeUnit
 import java.util.concurrent.atomic.{AtomicBoolean, AtomicLong}
 import java.util.concurrent.locks.ReentrantLock
 
@@ -40,7 +41,7 @@ import org.apache.spark.network.shuffle.mesos.MesosExternalShuffleClient
 import org.apache.spark.rpc.RpcEndpointAddress
 import org.apache.spark.scheduler.{SlaveLost, TaskSchedulerImpl}
 import org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend
-import org.apache.spark.util.Utils
+import org.apache.spark.util.{ThreadUtils, Utils}
 
 /**
  * A SchedulerBackend that runs tasks on Mesos, but uses "coarse-grained" tasks, where it holds
@@ -90,6 +91,7 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
 
   // Synchronization protected by stateLock
   private[this] var stopCalled: Boolean = false
+  private[this] var offersSuppressed: Boolean = false
 
   private val launcherBackend = new LauncherBackend() {
     override protected def conf: SparkConf = sc.conf
@@ -187,6 +189,10 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
   private val schedulerUuid: String = UUID.randomUUID().toString
   private val nextExecutorNumber = new AtomicLong()
 
+  private val reviveOffersExecutorService =
+    ThreadUtils.newDaemonSingleThreadScheduledExecutor("mesos-revive-thread")
+  private val reviveIntervalMs = conf.get(REVIVE_OFFERS_INTERVAL)
+
   override def start() {
     super.start()
 
@@ -218,6 +224,18 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
 
     launcherBackend.setState(SparkAppHandle.State.SUBMITTED)
     startScheduler(driver)
+
+    // Periodic check if there is a need to revive mesos offers
+    reviveOffersExecutorService.scheduleAtFixedRate(new Runnable {
+      override def run(): Unit = {
+        stateLock.synchronized {
+          if (!offersSuppressed) {
+            logDebug("scheduled mesos offers revive")
+            schedulerDriver.reviveOffers
+          }
+        }
+      }
+    }, reviveIntervalMs, reviveIntervalMs, TimeUnit.MILLISECONDS)
   }
 
   def createCommand(offer: Offer, numCores: Int, taskId: String): CommandInfo = {
@@ -360,6 +378,8 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
   override def resourceOffers(d: org.apache.mesos.SchedulerDriver, offers: JList[Offer]) {
     stateLock.synchronized {
       metricsSource.recordOffers(offers.size)
+      logInfo(s"Received ${offers.size} resource offers.")
+
       if (stopCalled) {
         logDebug("Ignoring offers during shutdown")
         // Driver should simply return a stopped status on race
@@ -370,9 +390,10 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
       }
 
       if (numExecutors >= executorLimit) {
-        logDebug("Executor limit reached. numExecutors: " + numExecutors +
-          " executorLimit: " + executorLimit)
         offers.asScala.map(_.getId).foreach(d.declineOffer)
+        logInfo("Executor limit reached. numExecutors: " + numExecutors +
+          " executorLimit: " + executorLimit + " . Suppressing further offers.")
+        suppressOffers(Option(d))
         launchingExecutors = false
         return
       } else {
@@ -382,8 +403,6 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
         }
       }
 
-      logDebug(s"Received ${offers.size} resource offers.")
-
       val (matchedOffers, unmatchedOffers) = offers.asScala.partition { offer =>
         val offerAttributes = toAttributeMap(offer.getAttributesList)
         matchesAttributeRequirements(slaveOfferConstraints, offerAttributes)
@@ -416,6 +435,7 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
   private def handleMatchedOffers(
       driver: org.apache.mesos.SchedulerDriver, offers: mutable.Buffer[Offer]): Unit = {
     val tasks = buildMesosTasks(offers)
+    var suppressionRequired = false
     for (offer <- offers) {
       val offerAttributes = toAttributeMap(offer.getAttributesList)
       val offerMem = getResource(offer.getResourcesList, "mem")
@@ -455,6 +475,7 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
           Collections.singleton(offer.getId),
           offerTasks.asJava)
       } else if (totalCoresAcquired >= maxCores) {
+        suppressionRequired = true
         // Reject an offer for a configurable amount of time to avoid starving other frameworks
         metricsSource.recordDeclineFinished
         declineOffer(driver,
@@ -469,6 +490,11 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
           Some("Offer was declined due to unmet task launch constraints."))
       }
     }
+
+    if (suppressionRequired) {
+      logInfo("Max core number is reached. Suppressing further offers.")
+      suppressOffers(Option.empty)
+    }
   }
 
   private def getContainerInfo(conf: SparkConf): ContainerInfo.Builder = {
@@ -696,18 +722,36 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
         }
         executorTerminated(d, slaveId, taskId, s"Executor finished with state $state")
         // In case we'd rejected everything before but have now lost a node
-        metricsSource.recordRevive
-        d.reviveOffers
+        if (state != TaskState.FINISHED) {
+          logInfo("Reviving offers due to a failed executor task.")
+          reviveOffers(Option(d))
+        }
       }
     }
   }
 
+  private def reviveOffers(driver: Option[org.apache.mesos.SchedulerDriver]): Unit = {
+    stateLock.synchronized {
+      metricsSource.recordRevive
+      offersSuppressed = false
+      driver.getOrElse(schedulerDriver).reviveOffers
+    }
+  }
+
+  private def suppressOffers(driver: Option[org.apache.mesos.SchedulerDriver]): Unit = {
+    stateLock.synchronized {
+      offersSuppressed = true
+      driver.getOrElse(schedulerDriver).suppressOffers
+    }
+  }
+
   override def error(d: org.apache.mesos.SchedulerDriver, message: String) {
     logError(s"Mesos error: $message")
     scheduler.error(message)
   }
 
   override def stop() {
+    reviveOffersExecutorService.shutdownNow()
     stopSchedulerBackend()
     launcherBackend.setState(SparkAppHandle.State.FINISHED)
     launcherBackend.close()
@@ -791,7 +835,12 @@ private[spark] class MesosCoarseGrainedSchedulerBackend(
     // We don't truly know if we can fulfill the full amount of executors
     // since at coarse grain it depends on the amount of slaves available.
     logInfo("Capping the total amount of executors to " + requestedTotal)
+    val reviveNeeded = executorLimit < requestedTotal
     executorLimitOption = Some(requestedTotal)
+    if (reviveNeeded && schedulerDriver != null) {
+      logInfo("The executor limit increased. Reviving offers.")
+      reviveOffers(Option.empty)
+    }
     // Update the locality wait start time to continue trying for locality.
     localityWaitStartTime = System.currentTimeMillis()
     true

resource-managers/mesos/src/test/scala/org/apache/spark/scheduler/cluster/mesos/MesosClusterSchedulerSuite.scala

@@ -593,6 +593,7 @@ class MesosClusterSchedulerSuite extends SparkFunSuite with LocalSparkContext wi
 
     // Offer new resource to retry driver on a new agent
     val agent2 = SlaveID.newBuilder().setValue("s2").build()
+    Thread.sleep(1500)
     scheduler.resourceOffers(driver, Collections.singletonList(offers(1)))
     taskStatus = TaskStatus.newBuilder()
       .setTaskId(TaskID.newBuilder().setValue(response.submissionId).build())

resource-managers/mesos/src/test/scala/org/apache/spark/scheduler/cluster/mesos/MesosCoarseGrainedSchedulerBackendSuite.scala

@@ -339,6 +339,90 @@ class MesosCoarseGrainedSchedulerBackendSuite extends SparkFunSuite
     assert(taskInfos.length == 2)
   }
 
+  test("scheduler backend suppresses mesos offers when max core count reached") {
+    val executorCores = 2
+    val executors = 2
+    setBackend(Map(
+      "spark.executor.cores" -> executorCores.toString(),
+      "spark.cores.max" -> (executorCores * executors).toString()))
+
+    val executorMemory = backend.executorMemory(sc)
+    offerResources(List(
+      Resources(executorMemory, executorCores),
+      Resources(executorMemory, executorCores),
+      Resources(executorMemory, executorCores)))
+
+    assert(backend.getTaskCount() == 2)
+    verify(driver, times(1)).suppressOffers()
+
+    // Finishing at least one task should trigger a revive
+    val status = createTaskStatus("0", "s1", TaskState.TASK_FAILED)
+    backend.statusUpdate(driver, status)
+    verify(driver, times(1)).reviveOffers()
+
+    offerResources(List(
+      Resources(executorMemory, executorCores)))
+    verify(driver, times(2)).suppressOffers()
+  }
+
+  test("scheduler backend suppresses mesos offers when the executor cap is reached") {
+    val executorCores = 1
+    val executors = 10
+    setBackend(Map(
+      "spark.executor.cores" -> executorCores.toString(),
+      "spark.cores.max" -> (executorCores * executors).toString()))
+
+    val executorMemory = backend.executorMemory(sc)
+    offerResources(List(
+      Resources(executorMemory, executorCores),
+      Resources(executorMemory, executorCores),
+      Resources(executorMemory, executorCores)))
+
+    assert(backend.getTaskCount() == 3)
+    verify(driver, times(0)).suppressOffers()
+
+    assert(backend.doRequestTotalExecutors(3).futureValue)
+    offerResources(List(
+      Resources(executorMemory, executorCores)))
+    verify(driver, times(1)).suppressOffers()
+
+    assert(backend.doRequestTotalExecutors(2).futureValue)
+    verify(driver, times(0)).reviveOffers()
+
+    // Finishing at least one task should trigger a revive
+    val status = createTaskStatus("0", "s1", TaskState.TASK_FAILED)
+    backend.statusUpdate(driver, status)
+    verify(driver, times(1)).reviveOffers()
+
+    offerResources(List(
+      Resources(executorMemory, executorCores)))
+    verify(driver, times(2)).suppressOffers()
+  }
+
+  test("scheduler periodically revives mesos offers if needed") {
+    val executorCores = 1
+    val executors = 3
+    setBackend(Map(
+      "spark.mesos.scheduler.revive.interval" -> "1s",
+      "spark.executor.cores" -> executorCores.toString(),
+      "spark.cores.max" -> (executorCores * executors).toString()))
+
+    val executorMemory = backend.executorMemory(sc)
+    offerResources(List(
+      Resources(executorMemory, executorCores)))
+
+    assert(backend.getTaskCount() == 1)
+    verify(driver, times(0)).suppressOffers()
+
+    // Verify that offers are revived every second
+    Thread.sleep(1500)
+    verify(driver, times(1)).reviveOffers()
+
+    Thread.sleep(1000)
+    verify(driver, times(2)).reviveOffers()
+
+  }
+
   test("mesos doesn't register twice with the same shuffle service") {
     setBackend(Map(SHUFFLE_SERVICE_ENABLED.key -> "true"))
     val (mem, cpu) = (backend.executorMemory(sc), 4)