Github user mccheah commented on a diff in the pull request:
https://github.com/apache/spark/pull/19468#discussion_r147003383
--- Diff:
resource-managers/kubernetes/core/src/main/scala/org/apache/spark/scheduler/cluster/k8s/ExecutorPodFactory.scala
---
@@ -0,0 +1,229 @@
+/*
+ * 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 scala.collection.JavaConverters._
+
+import io.fabric8.kubernetes.api.model._
+
+import org.apache.spark.{SparkConf, SparkException}
+import org.apache.spark.deploy.k8s.ConfigurationUtils
+import org.apache.spark.deploy.k8s.config._
+import org.apache.spark.deploy.k8s.constants._
+import org.apache.spark.util.Utils
+
+/**
+ * Configures executor pods. Construct one of these with a SparkConf to
set up properties that are
+ * common across all executors. Then, pass in dynamic parameters into
createExecutorPod.
+ */
+private[spark] trait ExecutorPodFactory {
+ def createExecutorPod(
+ executorId: String,
+ applicationId: String,
+ driverUrl: String,
+ executorEnvs: Seq[(String, String)],
+ driverPod: Pod,
+ nodeToLocalTaskCount: Map[String, Int]): Pod
+}
+
+private[spark] class ExecutorPodFactoryImpl(sparkConf: SparkConf)
+ extends ExecutorPodFactory {
+
+ import ExecutorPodFactoryImpl._
+
+ private val executorExtraClasspath = sparkConf.get(
+ org.apache.spark.internal.config.EXECUTOR_CLASS_PATH)
+ private val executorJarsDownloadDir =
sparkConf.get(INIT_CONTAINER_JARS_DOWNLOAD_LOCATION)
+
+ private val executorLabels =
ConfigurationUtils.parsePrefixedKeyValuePairs(
+ sparkConf,
+ KUBERNETES_EXECUTOR_LABEL_PREFIX,
+ "executor label")
+ require(
+ !executorLabels.contains(SPARK_APP_ID_LABEL),
+ s"Custom executor labels cannot contain $SPARK_APP_ID_LABEL as it is
reserved for Spark.")
+ require(
+ !executorLabels.contains(SPARK_EXECUTOR_ID_LABEL),
+ s"Custom executor labels cannot contain $SPARK_EXECUTOR_ID_LABEL as it
is reserved for" +
+ s" Spark.")
+
+ private val executorAnnotations =
+ ConfigurationUtils.parsePrefixedKeyValuePairs(
+ sparkConf,
+ KUBERNETES_EXECUTOR_ANNOTATION_PREFIX,
+ "executor annotation")
+ private val nodeSelector =
+ ConfigurationUtils.parsePrefixedKeyValuePairs(
+ sparkConf,
+ KUBERNETES_NODE_SELECTOR_PREFIX,
+ "node selector")
+
+ private val executorDockerImage = sparkConf.get(EXECUTOR_DOCKER_IMAGE)
+ private val dockerImagePullPolicy =
sparkConf.get(DOCKER_IMAGE_PULL_POLICY)
+ private val executorPort = sparkConf.getInt("spark.executor.port",
DEFAULT_STATIC_PORT)
+ private val blockmanagerPort = sparkConf
+ .getInt("spark.blockmanager.port", DEFAULT_BLOCKMANAGER_PORT)
+ private val kubernetesDriverPodName = sparkConf
+ .get(KUBERNETES_DRIVER_POD_NAME)
+ .getOrElse(throw new SparkException("Must specify the driver pod
name"))
+
+ private val executorPodNamePrefix =
sparkConf.get(KUBERNETES_EXECUTOR_POD_NAME_PREFIX)
+
+ private val executorMemoryMiB =
sparkConf.get(org.apache.spark.internal.config.EXECUTOR_MEMORY)
+ private val executorMemoryString = sparkConf.get(
+ org.apache.spark.internal.config.EXECUTOR_MEMORY.key,
+ org.apache.spark.internal.config.EXECUTOR_MEMORY.defaultValueString)
+
+ private val memoryOverheadMiB = sparkConf
+ .get(KUBERNETES_EXECUTOR_MEMORY_OVERHEAD)
+ .getOrElse(math.max((MEMORY_OVERHEAD_FACTOR * executorMemoryMiB).toInt,
+ MEMORY_OVERHEAD_MIN_MIB))
+ private val executorMemoryWithOverhead = executorMemoryMiB +
memoryOverheadMiB
+
+ private val executorCores = sparkConf.getDouble("spark.executor.cores",
1d)
+ private val executorLimitCores =
sparkConf.getOption(KUBERNETES_EXECUTOR_LIMIT_CORES.key)
+
+ override def createExecutorPod(
+ executorId: String,
+ applicationId: String,
+ driverUrl: String,
+ executorEnvs: Seq[(String, String)],
+ driverPod: Pod,
+ nodeToLocalTaskCount: Map[String, Int]): Pod = {
+ val name = s"$executorPodNamePrefix-exec-$executorId"
+
+ // hostname must be no longer than 63 characters, so take the last 63
characters of the pod
+ // name as the hostname. This preserves uniqueness since the end of
name contains
+ // executorId and applicationId
+ val hostname = name.substring(Math.max(0, name.length - 63))
+ val resolvedExecutorLabels = Map(
+ SPARK_EXECUTOR_ID_LABEL -> executorId,
+ SPARK_APP_ID_LABEL -> applicationId,
+ SPARK_ROLE_LABEL -> SPARK_POD_EXECUTOR_ROLE) ++
+ executorLabels
+ val executorMemoryQuantity = new QuantityBuilder(false)
+ .withAmount(s"${executorMemoryMiB}Mi")
+ .build()
+ val executorMemoryLimitQuantity = new QuantityBuilder(false)
+ .withAmount(s"${executorMemoryWithOverhead}Mi")
+ .build()
+ val executorCpuQuantity = new QuantityBuilder(false)
+ .withAmount(executorCores.toString)
+ .build()
+ val executorExtraClasspathEnv = executorExtraClasspath.map { cp =>
+ new EnvVarBuilder()
+ .withName(ENV_EXECUTOR_EXTRA_CLASSPATH)
+ .withValue(cp)
+ .build()
+ }
+ val executorExtraJavaOptionsEnv = sparkConf
+ .get(org.apache.spark.internal.config.EXECUTOR_JAVA_OPTIONS)
+ .map { opts =>
+ val delimitedOpts = Utils.splitCommandString(opts)
+ delimitedOpts.zipWithIndex.map {
+ case (opt, index) =>
+ new
EnvVarBuilder().withName(s"$ENV_JAVA_OPT_PREFIX$index").withValue(opt).build()
+ }
+ }.getOrElse(Seq.empty[EnvVar])
+ val executorEnv = (Seq(
+ (ENV_EXECUTOR_PORT, executorPort.toString),
+ (ENV_DRIVER_URL, driverUrl),
+ // Executor backend expects integral value for executor cores, so
round it up to an int.
+ (ENV_EXECUTOR_CORES, math.ceil(executorCores).toInt.toString),
+ (ENV_EXECUTOR_MEMORY, executorMemoryString),
+ (ENV_APPLICATION_ID, applicationId),
+ (ENV_EXECUTOR_ID, executorId),
+ (ENV_MOUNTED_CLASSPATH, s"$executorJarsDownloadDir/*")) ++
executorEnvs)
+ .map(env => new EnvVarBuilder()
+ .withName(env._1)
+ .withValue(env._2)
+ .build()
+ ) ++ Seq(
+ new EnvVarBuilder()
+ .withName(ENV_EXECUTOR_POD_IP)
+ .withValueFrom(new EnvVarSourceBuilder()
+ .withNewFieldRef("v1", "status.podIP")
+ .build())
+ .build()
+ ) ++ executorExtraJavaOptionsEnv ++ executorExtraClasspathEnv.toSeq
+ val requiredPorts = Seq(
+ (EXECUTOR_PORT_NAME, executorPort),
+ (BLOCK_MANAGER_PORT_NAME, blockmanagerPort))
+ .map(port => {
+ new ContainerPortBuilder()
+ .withName(port._1)
+ .withContainerPort(port._2)
+ .build()
+ })
+
+ val executorContainer = new ContainerBuilder()
+ .withName(s"executor")
+ .withImage(executorDockerImage)
+ .withImagePullPolicy(dockerImagePullPolicy)
+ .withNewResources()
+ .addToRequests("memory", executorMemoryQuantity)
+ .addToLimits("memory", executorMemoryLimitQuantity)
+ .addToRequests("cpu", executorCpuQuantity)
+ .endResources()
+ .addAllToEnv(executorEnv.asJava)
+ .withPorts(requiredPorts.asJava)
+ .build()
+
+ val executorPod = new PodBuilder()
+ .withNewMetadata()
+ .withName(name)
+ .withLabels(resolvedExecutorLabels.asJava)
+ .withAnnotations(executorAnnotations.asJava)
+ .withOwnerReferences()
+ .addNewOwnerReference()
+ .withController(true)
+ .withApiVersion(driverPod.getApiVersion)
+ .withKind(driverPod.getKind)
+ .withName(driverPod.getMetadata.getName)
+ .withUid(driverPod.getMetadata.getUid)
+ .endOwnerReference()
+ .endMetadata()
+ .withNewSpec()
+ .withHostname(hostname)
+ .withRestartPolicy("Never")
--- End diff --
> Am I correct in assuming that each executor will download all the
artifacts (jars, files, etc) even if there are others colocated on the same
node ?
This will eventually be correct when we introduce the init-container that
fetches dependencies before drivers and executors launch.
The notion of locality is a tricky one in the context of Kubernetes. Since
Spark drivers and executors are running in a containerized setting, each
process is running in a completely isolated setting. This means that multiple
pods that even run on the same kubelet will not have any shared storage. This
creates some interesting problems that we will address in the future, including
that of needing to localize dependencies.
We expect a common workflow for Spark on Kubernetes for users to install
dependencies into their driver/executor container's images. This allows one to
bypass the overhead of localizing dependencies from remote locations
repeatedly. This is especially true if a workload is run multiple times on the
same cluster, because the image binary will be cached on the Kubelets in
between multiple executions.
But we're also introducing support for Spark applications to be able to
localize dependencies from remote locations, such as an HTTP file server or
HDFS. That will come in a future commit.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
