Github user mccheah commented on a diff in the pull request:
https://github.com/apache/spark/pull/19468#discussion_r147024086
--- 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()
--- End diff --
@mridulm In response to each of your four points:
- IPv6: IP addresses are strictly managed by the Kubernetes framework, so
it's unlikely we're going to run into differences between Ipv4 and Ipv6 in
different Kubernetes clusters. We should assume that one of these two address
types are being used across all clusters and work with that. So far I've only
seen Kubernetes assign IPv4 addresses.
- No support for multihomed machines of multi-routable IP's: Again, since
Kubernetes is managing the IP address and routability of pods, we can
understand what the framework will do and work with that. I don't think the
framework does anything fancy in this space, but maybe @foxish or others have
ideas?
- Distributed stores: We've thought about this and have some work done on
our fork in regards to this that we will eventually contribute back here.
@kimoonkim has done some work on this. The short version of the situation we've
had to work around is that the container that runs the Spark processes now has
a different IP address from the HDFS node that might be colocated with it
physically.
- Rack awareness - @kimoonkim has also done work on this, and similar
concerns to the point above have come up.
As a general note, locality is non-trivial in Kubernetes because no two
pods will ever share the same IP address, and pods do not share the same IP
address as the physical host that is running it. The Kubernetes code needs to
be intelligent about knowing which pods are co-located on the same underlying
Kubelet. And of course, it's reasonable to believe that the above four
considerations are not exhaustive, but we'll address unforeseen factors as they
come up.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
