squito commented on a change in pull request #24704: [SPARK-20286][core] Improve logic for timing out executors in dynamic allocation. URL: https://github.com/apache/spark/pull/24704#discussion_r289084451
########## File path: core/src/main/scala/org/apache/spark/scheduler/dynalloc/ExecutorMonitor.scala ########## @@ -0,0 +1,293 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.spark.scheduler.dynalloc + +import java.util.concurrent.{ConcurrentHashMap, TimeUnit} +import java.util.concurrent.atomic.AtomicLong + +import scala.collection.JavaConverters._ +import scala.collection.mutable + +import org.apache.spark._ +import org.apache.spark.internal.Logging +import org.apache.spark.internal.config._ +import org.apache.spark.scheduler._ +import org.apache.spark.storage.RDDBlockId +import org.apache.spark.util.Clock + +/** + * A monitor for executor activity, used by ExecutorAllocationManager to detect idle executors. + */ +private[spark] class ExecutorMonitor( + conf: SparkConf, + client: ExecutorAllocationClient, + clock: Clock) extends SparkListener with Logging { + + private val idleTimeoutMs = TimeUnit.SECONDS.toMillis( + conf.get(DYN_ALLOCATION_EXECUTOR_IDLE_TIMEOUT)) + private val storageTimeoutMs = TimeUnit.SECONDS.toMillis( + conf.get(DYN_ALLOCATION_CACHED_EXECUTOR_IDLE_TIMEOUT)) + private val fetchFromShuffleSvcEnabled = conf.get(SHUFFLE_SERVICE_ENABLED) && + conf.get(SHUFFLE_SERVICE_FETCH_RDD_ENABLED) + + private val executors = new ConcurrentHashMap[String, Tracker]() + + // The following fields are an optimization to avoid having to scan all executors on every EAM + // schedule interval to find out which ones are timed out. They keep track of when the next + // executor timeout is expected to happen, and the current list of timed out executors. There's + // also a flag that forces the EAM task to recompute the timed out executors, in case some event + // arrives on the listener bus that may cause the current list of timed out executors to change. + // + // There's also per-executor state used for this purpose, so that recomputations can be triggered + // only when really necessary. + // + // Note that this isn't meant to, and cannot, always make the right decision about which executors + // are indeed timed out. For example, the EAM thread may detect a timed out executor while a new + // "task start" event has just been posted to the listener bus and hasn't yet been delivered to + // this listener. There are safeguards in other parts of the code that would prevent that executor + // from being removed. + private var nextTimeout = new AtomicLong(Long.MaxValue) + private var timedOutExecs = Seq.empty[String] + + def reset(): Unit = { + executors.clear() + nextTimeout.set(Long.MaxValue) + timedOutExecs = Nil + } + + def timedOutExecutors(): Seq[String] = { + val now = clock.getTimeMillis() + if (now >= nextTimeout.get()) { + // Temporarily set the next timeout at Long.MaxValue. This ensures that after + // scanning all executors below, we know when the next timeout for non-timed out + // executors is (whether that update came from the scan, or from a new event + // arriving in a different thread). + nextTimeout.set(Long.MaxValue) + + var newNextTimeout = Long.MaxValue + timedOutExecs = executors.asScala + .filter { case (_, exec) => !exec.pendingRemoval } + .filter { case (_, exec) => + val deadline = exec.timeoutAt + if (deadline > now) { + newNextTimeout = math.min(newNextTimeout, deadline) + exec.timedOut = false + false + } else { + exec.timedOut = true + + // An event arriving while this scan is happening may cause the deadline for + // the executor to change after it was read above. Check the deadline again, + // and if it changed, don't consider this executor for removal yet. + val newDeadline = exec.timeoutAt + if (newDeadline > now) { + exec.timedOut = false + newNextTimeout = math.min(newNextTimeout, newDeadline) + false + } else { + true + } + } + } + .keys + .toSeq + updateNextTimeout(newNextTimeout) + } + timedOutExecs + } + + /** + * Mark the given executors as pending to be removed. Should only be called in the EAM thread. + */ + def executorsKilled(ids: Seq[String]): Unit = { + ids.foreach { id => + val tracker = executors.get(id) + if (tracker != null) { + tracker.pendingRemoval = 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 pendingRemovalCount: Int = executors.asScala.count { case (_, exec) => exec.pendingRemoval } + + override def onTaskStart(event: SparkListenerTaskStart): Unit = { + val executorId = event.taskInfo.executorId + // Guard against a late arriving task start event (SPARK-26927). + if (client.isExecutorActive(executorId)) { + val exec = ensureExecutorIsTracked(executorId) + exec.updateRunningTasks(1) + } + } + + override def onTaskEnd(event: SparkListenerTaskEnd): Unit = { + val executorId = event.taskInfo.executorId + val exec = executors.get(executorId) + if (exec != null) { + exec.updateRunningTasks(-1) + } + } + + override def onExecutorAdded(event: SparkListenerExecutorAdded): Unit = { + val exec = ensureExecutorIsTracked(event.executorId) + exec.updateRunningTasks(0) + logInfo(s"New executor ${event.executorId} has registered (new total is ${executors.size()})") + } + + override def onExecutorRemoved(event: SparkListenerExecutorRemoved): Unit = { + val removed = executors.remove(event.executorId) + if (removed != null) { + logInfo(s"Executor ${event.executorId} removed (new total is ${executors.size()})") + if (!removed.pendingRemoval) { + nextTimeout.set(Long.MinValue) + } + } + } + + override def onBlockUpdated(event: SparkListenerBlockUpdated): Unit = { + if (!event.blockUpdatedInfo.blockId.isInstanceOf[RDDBlockId]) { + return + } + + val exec = ensureExecutorIsTracked(event.blockUpdatedInfo.blockManagerId.executorId) + val storageLevel = event.blockUpdatedInfo.storageLevel + val blockId = event.blockUpdatedInfo.blockId.asInstanceOf[RDDBlockId] + + // SPARK-27677. When a block can be fetched from the external shuffle service, the executor can + // be removed without hurting the application too much, since the cached data is still + // available. So don't count blocks that can be served by the external service. + if (storageLevel.isValid && (!fetchFromShuffleSvcEnabled || !storageLevel.useDisk)) { + val hadCachedBlocks = exec.cachedBlocks.nonEmpty + + val blocks = exec.cachedBlocks.getOrElseUpdate(blockId.rddId, + new mutable.BitSet(blockId.splitIndex)) + blocks += blockId.splitIndex + + if (!hadCachedBlocks) { + exec.updateTimeout() + } + } else { + exec.cachedBlocks.get(blockId.rddId).foreach { blocks => + blocks -= blockId.splitIndex + if (blocks.isEmpty) { + exec.cachedBlocks -= blockId.rddId + if (exec.cachedBlocks.isEmpty) { + exec.updateTimeout() + } + } + } + } + } + + override def onUnpersistRDD(event: SparkListenerUnpersistRDD): Unit = { + executors.values().asScala.foreach { exec => + exec.cachedBlocks -= event.rddId + if (exec.cachedBlocks.isEmpty) { + exec.updateTimeout() + } + } + } + + // Visible for testing. + private[dynalloc] def isExecutorIdle(id: String): Boolean = { + Option(executors.get(id)).map(_.isIdle).getOrElse(throw new NoSuchElementException(id)) + } + + // Visible for testing + private[dynalloc] def timedOutExecutors(when: Long): Seq[String] = { + executors.asScala.flatMap { case (id, tracker) => + if (tracker.timeoutAt <= when) Some(id) else None + }.toSeq + } + + // Visible for testing + def executorsPendingToRemove(): Set[String] = { + executors.asScala.filter { case (_, exec) => exec.pendingRemoval }.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` + * event, which is possible because these events are posted in different threads. (see SPARK-4951) + */ + private def ensureExecutorIsTracked(id: String): Tracker = { + executors.computeIfAbsent(id, _ => new Tracker()) + } + + private def updateNextTimeout(newValue: Long): Unit = { + while (true) { + val current = nextTimeout.get() + if (newValue >= current || nextTimeout.compareAndSet(current, newValue)) { + return + } + } + } + + private class Tracker { + @volatile var timeoutAt: Long = Long.MaxValue + + // Tracks whether this executor is thought to be timed out. It's used to detect when the list + // of timed out executors needs to be updated due to the executor's state changing. + @volatile var timedOut: Boolean = false + + var pendingRemoval: Boolean = false + + private var idleStart: Long = -1 + private var runningTasks: Int = 0 + + // Maps RDD IDs to the partition IDs stored in the executor. + // This should only be used in the event thread. + val cachedBlocks = new mutable.HashMap[Int, mutable.BitSet]() + + // For testing. + def isIdle: Boolean = idleStart >= 0 + + def updateRunningTasks(delta: Int): Unit = { + runningTasks = math.max(0, runningTasks + delta) + idleStart = if (runningTasks == 0) clock.getTimeMillis() else -1L + updateTimeout() + } + + def updateTimeout(): Unit = { + val oldDeadline = timeoutAt + val newDeadline = if (idleStart >= 0) { + idleStart + (if (cachedBlocks.nonEmpty) storageTimeoutMs else idleTimeoutMs) + } else { + Long.MaxValue + } + + // Update the timeout before checking whether a recomputation is needed. This ensures + // that if the EAM sees the new timeout when double-checking whether an executor is really + // timed out, after setting the "timedOut" flag checked below. + timeoutAt = newDeadline + + // If the executor was thought to be timed out, but the new deadline is later than the + // old one, ask the EAM thread to update the list of timed out executors. + if (timedOut && newDeadline > oldDeadline) { + nextTimeout.set(Long.MinValue) Review comment: > it gets the list of timed out executors and asks cluster manager to kill them by ExecutorAllocationClient immediately. So, if an executor has already been considered timed out and killing after last round, will there be any difference for that executor if we eagerly update the list of timed out executors in next round ? only the CoarseGrainedSchedulerBackend has a consistent view of which executors are actually running tasks, and so it has ultimate say in whether or not to kill an executor. You might not have killed an executor earlier because it was actually busy. So here is where you "catch up" with the view of the CGSB, so you know that executor shouldn't be considered timed out and requested to be killed next time around. (at least I think so, marcelo can confirm ...) ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org
