Dynamic Executor Scaling

conf/kubernetes-shuffle-service.yaml

@@ -0,0 +1,36 @@
+apiVersion: extensions/v1beta1
+kind: DaemonSet
+metadata:
+  labels:
+    app: spark-shuffle-service
+    spark-version: 2.1.0
+  name: shuffle
+spec:
+  template:
+    metadata:
+      labels:
+        app: spark-shuffle-service
+        spark-version: 2.1.0
+    spec:
+      volumes:
+        - name: temp-volume
+          hostPath:
+            path: '/var/tmp' # change this path according to your cluster configuration.
+      containers:
+        - name: shuffle
+          # This is an official image that is built
+          # from the dockerfiles/shuffle directory
+          # in the spark distribution.
+          image: kubespark/spark-shuffle:v2.1.0-kubernetes-0.1.0-alpha.3
+          volumeMounts:
+            - mountPath: '/tmp'
+              name: temp-volume
+              # more volumes can be mounted here.
+              # The spark job must be configured to use these
+              # mounts using the configuration:
+              #   spark.kubernetes.shuffle.dir=<mount-1>,<mount-2>,...
+          resources:
+             requests:
+               cpu: "1"
+             limits:
+               cpu: "1"

core/src/main/scala/org/apache/spark/executor/CoarseGrainedExecutorBackend.scala

@@ -200,7 +200,7 @@ private[spark] object CoarseGrainedExecutorBackend extends Logging {
         new SecurityManager(executorConf),
         clientMode = true)
       val driver = fetcher.setupEndpointRefByURI(driverUrl)
-      val cfg = driver.askWithRetry[SparkAppConfig](RetrieveSparkAppConfig)
+      val cfg = driver.askWithRetry[SparkAppConfig](RetrieveSparkAppConfig(executorId))
       val props = cfg.sparkProperties ++ Seq[(String, String)](("spark.app.id", appId))
       fetcher.shutdown()
 

core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedClusterMessage.scala

@@ -28,7 +28,7 @@ private[spark] sealed trait CoarseGrainedClusterMessage extends Serializable
 
 private[spark] object CoarseGrainedClusterMessages {
 
-  case object RetrieveSparkAppConfig extends CoarseGrainedClusterMessage
+  case class RetrieveSparkAppConfig(executorId: String) extends CoarseGrainedClusterMessage
 
   case class SparkAppConfig(
       sparkProperties: Seq[(String, String)],

core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala

@@ -206,7 +206,7 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val rpcEnv: Rp
         removeExecutor(executorId, reason)
         context.reply(true)
 
-      case RetrieveSparkAppConfig =>
+      case RetrieveSparkAppConfig(executorId) =>
         val reply = SparkAppConfig(sparkProperties,
           SparkEnv.get.securityManager.getIOEncryptionKey())
         context.reply(reply)

core/src/main/scala/org/apache/spark/storage/BlockManager.scala

@@ -184,8 +184,14 @@ private[spark] class BlockManager(
     blockManagerId = if (idFromMaster != null) idFromMaster else id
 
     shuffleServerId = if (externalShuffleServiceEnabled) {
-      logInfo(s"external shuffle service port = $externalShuffleServicePort")
-      BlockManagerId(executorId, blockTransferService.hostName, externalShuffleServicePort)
+      val shuffleServerHostName = if (blockManagerId.isDriver) {
+        blockTransferService.hostName
+      } else {
+        conf.get("spark.shuffle.service.host", blockTransferService.hostName)
+      }
+      logInfo(s"external shuffle service host = $shuffleServerHostName, " +
+        s"port = $externalShuffleServicePort")
+      BlockManagerId(executorId, shuffleServerHostName, externalShuffleServicePort)
     } else {
       blockManagerId
     }

docs/running-on-kubernetes.md

@@ -189,7 +189,7 @@ from the other deployment modes. See the [configuration page](configuration.html
   <td>
     The namespace that will be used for running the driver and executor pods. When using
     <code>spark-submit</code> in cluster mode, this can also be passed to <code>spark-submit</code> via the
-    <code>--kubernetes-namespace</code> command line argument. The namespace must already exist.
+    <code>--kubernetes-namespace</code> command line argument.
   </td>
 </tr>
 <tr>
@@ -208,6 +208,30 @@ from the other deployment modes. See the [configuration page](configuration.html
     <a href="https://docs.docker.com/engine/reference/commandline/tag/">Docker tag</a> format.
   </td>
 </tr>
+<tr>
+  <td><code>spark.kubernetes.shuffle.namespace</code></td>
+  <td><code>default</code></td>
+  <td>
+    Namespace in which the shuffle service pods are present. The shuffle service must be
+    created in the cluster prior to attempts to use it.
+  </td>
+</tr>
+<tr>
+  <td><code>spark.kubernetes.shuffle.labels</code></td>
+  <td><code>(none)</code></td>
+  <td>
+    Labels that will be used to look up shuffle service pods. This should be a comma-separated list of label key-value pairs,
+    where each label is in the format <code>key=value</code>. The labels chosen must be such that
+    they match exactly one shuffle service pod on each node that executors are launched.
+  </td>
+</tr>
+<tr>
+  <td><code>spark.kubernetes.dynamic.allocation.size</code></td>
+  <td><code>5</code></td>
+  <td>
+    Number of executor pods to launch at once in each round of dynamic allocation.
+  </td>
+</tr>
 <tr>
   <td><code>spark.kubernetes.authenticate.submission.caCertFile</code></td>
   <td>(none)</td>
@@ -382,10 +406,41 @@ from the other deployment modes. See the [configuration page](configuration.html
 </tr>
 </table>
 
+## Dynamic Executor Scaling
+
+Spark on Kubernetes supports Dynamic Allocation with cluster mode. This mode requires running
+an external shuffle service. This is typically a [daemonset](https://kubernetes.io/docs/concepts/workloads/controllers/daemonset/)
+with a provisioned [hostpath](https://kubernetes.io/docs/concepts/storage/volumes/#hostpath) volume.
+This shuffle service may be shared by executors belonging to different SparkJobs. Using Spark with dynamic allocation
+on Kubernetes assumes that a cluster administrator has set up one or more shuffle-service daemonsets in the cluster.
+
+A sample configuration file is provided in `conf/kubernetes-shuffle-service.yaml` which can be customized as needed
+for a particular cluster. It is important to note that `spec.template.metadata.labels` are setup appropriately for the shuffle
+service because there may be multiple shuffle service instances running in a cluster. The labels give us a way to target a particular
+shuffle service.
+
+For example, if the shuffle service we want to use is in the default namespace, and
+has pods with labels `app=spark-shuffle-service` and `spark-version=2.1.0`, we can
+use those tags to target that particular shuffle service at job launch time. In order to run a job with dynamic allocation enabled,
+the command may then look like the following:
+
+    bin/spark-submit \
+      --deploy-mode cluster \
+      --class org.apache.spark.examples.GroupByTest \
+      --master k8s://<k8s-master>:<port> \
+      --kubernetes-namespace default \
+      --conf spark.app.name=group-by-test \
+      --conf spark.kubernetes.driver.docker.image=kubespark/spark-driver:latest \
+      --conf spark.kubernetes.executor.docker.image=kubespark/spark-executor:latest \
+      --conf spark.dynamicAllocation.enabled=true \
+      --conf spark.shuffle.service.enabled=true \
+      --conf spark.kubernetes.shuffle.namespace=default \
+      --conf spark.kubernetes.shuffle.labels="app=spark-shuffle-service,spark-version=2.1.0" \
+      examples/jars/spark_examples_2.11-2.2.0.jar 10 400000 2
+
 ## Current Limitations
 
 Running Spark on Kubernetes is currently an experimental feature. Some restrictions on the current implementation that
 should be lifted in the future include:
-* Applications can only use a fixed number of executors. Dynamic allocation is not supported.
 * Applications can only run in cluster mode.
 * Only Scala and Java applications can be run.

resource-managers/kubernetes/README.md

@@ -14,10 +14,10 @@ important matters to keep in mind when developing this feature.
 
 # Building Spark with Kubernetes Support
 
-To build Spark with Kubernetes support, use the `kubernetes` profile when invoking Maven.
+To build Spark with Kubernetes support, use the `kubernetes` profile when invoking Maven. For example, to simply compile
+the Kubernetes core implementation module along with its dependencies:
 
-    git checkout branch-2.1-kubernetes
-    build/mvn package -Pkubernetes -DskipTests
+    build/mvn compile -Pkubernetes -pl resource-managers/kubernetes/core -am -DskipTests
 
 To build a distribution of Spark with Kubernetes support, use the `dev/make-distribution.sh` script, and add the
 `kubernetes` profile as part of the build arguments. Any other build arguments can be specified as one would expect when

resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/kubernetes/Util.scala

@@ -0,0 +1,41 @@
+/*
+ * 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.deploy.kubernetes
+
+import org.apache.spark.SparkException
+
+object Util {
+  def parseKeyValuePairs(
+    maybeKeyValues: Option[String],
+    configKey: String,
+    keyValueType: String): Map[String, String] = {
+
+    maybeKeyValues.map(keyValues => {
+      keyValues.split(",").map(_.trim).filterNot(_.isEmpty).map(keyValue => {
+        keyValue.split("=", 2).toSeq match {
+          case Seq(k, v) =>
+            (k, v)
+          case _ =>
+            throw new SparkException(s"Custom $keyValueType set by $configKey must be a" +
+              s" comma-separated list of key-value pairs, with format <key>=<value>." +
+              s" Got value: $keyValue. All values: $keyValues")
+        }
+      }).toMap
+    }).getOrElse(Map.empty[String, String])
+  }
+}

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

@@ -157,6 +157,13 @@ package object config extends Logging {
       .stringConf
       .createOptional
 
+  private[spark] val SPARK_SHUFFLE_SERVICE_HOST =
+    ConfigBuilder("spark.shuffle.service.host")
+      .doc("Host for Spark Shuffle Service")
+      .internal()
+      .stringConf
+      .createOptional
+
   // Note that while we set a default for this when we start up the
   // scheduler, the specific default value is dynamically determined
   // based on the executor memory.
@@ -271,6 +278,38 @@ package object config extends Logging {
       .stringConf
       .createOptional
 
+  private[spark] val KUBERNETES_SHUFFLE_NAMESPACE =
+    ConfigBuilder("spark.kubernetes.shuffle.namespace")
+      .doc("Namespace of the shuffle service")
+      .stringConf
+      .createWithDefault("default")
+
+  private[spark] val KUBERNETES_SHUFFLE_SVC_IP =
+    ConfigBuilder("spark.kubernetes.shuffle.ip")
+      .doc("This setting is for debugging only. Setting this " +
+        "allows overriding the IP that the executor thinks its colocated " +
+        "shuffle service is on")
+      .stringConf
+      .createOptional
+
+  private[spark] val KUBERNETES_SHUFFLE_LABELS =
+    ConfigBuilder("spark.kubernetes.shuffle.labels")
+      .doc("Labels to identify the shuffle service")
+      .stringConf
+      .createOptional
+
+  private[spark] val KUBERNETES_SHUFFLE_DIR =
+    ConfigBuilder("spark.kubernetes.shuffle.dir")
+      .doc("Path to the shared shuffle directories.")
+      .stringConf
+      .createOptional
+
+  private[spark] val KUBERNETES_DYNAMIC_ALLOCATION_SIZE =
+    ConfigBuilder("spark.kubernetes.dynamic.allocation.size")
+      .doc("Number of pods to launch at once in each round of dynamic allocation.")
+      .intConf
+      .createWithDefault(5)
+
   private[spark] val DRIVER_SERVICE_MANAGER_TYPE =
     ConfigBuilder("spark.kubernetes.driver.serviceManagerType")
       .doc("A tag indicating which class to use for creating the Kubernetes service and" +

resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/kubernetes/constants.scala

@@ -113,4 +113,5 @@ package object constants {
     s"$INIT_CONTAINER_PROPERTIES_FILE_MOUNT_PATH/$INIT_CONTAINER_PROPERTIES_FILE_NAME"
   private[spark] val DOWNLOAD_JARS_VOLUME_NAME = "download-jars"
   private[spark] val DOWNLOAD_FILES_VOLUME_NAME = "download-files"
+  private[spark] val DEFAULT_SHUFFLE_MOUNT_NAME = "shuffle"
 }