Github user skonto commented on a diff in the pull request: https://github.com/apache/spark/pull/21366#discussion_r195350385 --- Diff: resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodsAllocator.scala --- @@ -0,0 +1,141 @@ +/* + * 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.cluster.k8s + +import java.util.concurrent.atomic.{AtomicInteger, AtomicLong} + +import io.fabric8.kubernetes.api.model.PodBuilder +import io.fabric8.kubernetes.client.KubernetesClient +import scala.collection.mutable + +import org.apache.spark.{SparkConf, SparkException} +import org.apache.spark.deploy.k8s.Config._ +import org.apache.spark.deploy.k8s.Constants._ +import org.apache.spark.deploy.k8s.KubernetesConf +import org.apache.spark.internal.Logging +import org.apache.spark.util.{Clock, Utils} + +private[spark] class ExecutorPodsAllocator( + conf: SparkConf, + executorBuilder: KubernetesExecutorBuilder, + kubernetesClient: KubernetesClient, + snapshotsStore: ExecutorPodsSnapshotsStore, + clock: Clock) extends Logging { + + private val EXECUTOR_ID_COUNTER = new AtomicLong(0L) + + private val totalExpectedExecutors = new AtomicInteger(0) + + private val podAllocationSize = conf.get(KUBERNETES_ALLOCATION_BATCH_SIZE) + + private val podAllocationDelay = conf.get(KUBERNETES_ALLOCATION_BATCH_DELAY) + + private val podCreationTimeout = math.max(podAllocationDelay * 5, 60000) + + private val kubernetesDriverPodName = conf + .get(KUBERNETES_DRIVER_POD_NAME) + .getOrElse(throw new SparkException("Must specify the driver pod name")) + + private val driverPod = kubernetesClient.pods() + .withName(kubernetesDriverPodName) + .get() + + // Executor IDs that have been requested from Kubernetes but have not been detected in any + // snapshot yet. Mapped to the timestamp when they were created. + private val newlyCreatedExecutors = mutable.Map.empty[Long, Long] + + def start(applicationId: String): Unit = { + snapshotsStore.addSubscriber(podAllocationDelay) { + onNewSnapshots(applicationId, _) + } + } + + def setTotalExpectedExecutors(total: Int): Unit = totalExpectedExecutors.set(total) + + private def onNewSnapshots(applicationId: String, snapshots: Seq[ExecutorPodsSnapshot]): Unit = { + newlyCreatedExecutors --= snapshots.flatMap(_.executorPods.keys) + // For all executors we've created against the API but have not seen in a snapshot + // yet - check the current time. If the current time has exceeded some threshold, + // assume that the pod was either never created (the API server never properly + // handled the creation request), or the API server created the pod but we missed + // both the creation and deletion events. In either case, delete the missing pod + // if possible, and mark such a pod to be rescheduled below. + newlyCreatedExecutors.foreach { case (execId, timeCreated) => + if (clock.getTimeMillis() - timeCreated > podCreationTimeout) { + logWarning(s"Executor with id $execId was not detected in the Kubernetes" + + s" cluster after $podCreationTimeout milliseconds despite the fact that a" + + " previous allocation attempt tried to create it. The executor may have been" + + " deleted but the application missed the deletion event.") + Utils.tryLogNonFatalError { + kubernetesClient + .pods() + .withLabel(SPARK_EXECUTOR_ID_LABEL, execId.toString) + .delete() --- End diff -- Shouldn't deleteFromSpark called here as well? Couldn't be the case that the executor exists at a higher level but K8s backend missed it?
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org