Scala Spark Docker Example

This example includes:

• Two Spark/Scala ML example docker containers and sample data
  • Can be run locally (instructions below)
  • Can be run on AWS EKS (Kubernetes) & S3 (no EMR/HDFS needed)
    • AWS-Setup-Guide-Spark-EKS.md- lists setup steps for Spark on AWS EKS
    • NOTE: we used the 'kops' service for this example, as it was required by EKS at the time we wrote this example.

---

Updated Example Code

Derived from these sources:

1. SGD Linear Regression Example with Apache Spark by Walker Rowe published May 23, 2017.
2. The example shows a linear regression example and has been modified to run as an app rather than in the interactive shell. Update, new example: Linear Regression from Spark Documentation. The new example has been updated to add serialization/deserialization and a split between training and test data.
3. Further reference Predicting Breast Cancer Using Apache Spark Machine Learning Logistic Regression by Carol McDonald published October 17, 2016.

---

How to Run Locally

This is an sbt project. Assuming we have a working scala and sbt, then execute sbt run from the project root. In addition to the log4j output from spark, you should also see a few lines of output from our example:

18/04/17 15:15:56 INFO TaskSetManager: Finished task 0.0 in stage 14.0 (TID 17) in 21 ms on localhost (executor driver) (1/1)
18/04/17 15:15:56 INFO TaskSchedulerImpl: Removed TaskSet 14.0, whose tasks have all completed, from pool
18/04/17 15:15:56 INFO DAGScheduler: ResultStage 14 (show at Main.scala:55) finished in 0.022 s
18/04/17 15:15:56 INFO DAGScheduler: Job 12 finished: show at Main.scala:55, took 0.026455 s
18/04/17 15:15:56 INFO CodeGenerator: Code generated in 8.628902 ms
+-------------------+--------------------+-------------------+
|              label|            features|         prediction|
+-------------------+--------------------+-------------------+
|-28.571478869743427|(10,[0,1,2,3,4,5,...|-1.5332357772511678|
|-26.736207182601724|(10,[0,1,2,3,4,5,...|-3.1990639907463776|
|-22.949825936196074|(10,[0,1,2,3,4,5,...|  2.068559275392233|
|-20.212077258958672|(10,[0,1,2,3,4,5,...| 0.5963989456221626|
|-17.026492264209548|(10,[0,1,2,3,4,5,...|-0.7387387189956682|
|-15.348871155379253|(10,[0,1,2,3,4,5,...|  -1.98575929759793|
|-13.039928064104615|(10,[0,1,2,3,4,5,...| 0.5942050121612523|
| -12.92222310337042|(10,[0,1,2,3,4,5,...|  2.203905559769596|
|-12.773226999251197|(10,[0,1,2,3,4,5,...| -2.736222698097398|
|-12.558575788856189|(10,[0,1,2,3,4,5,...|0.10007973294293643|
|-12.479280211451497|(10,[0,1,2,3,4,5,...|-0.9022515201372355|
| -12.46765638103286|(10,[0,1,2,3,4,5,...|-1.4621820914334354|
|-11.904986902675114|(10,[0,1,2,3,4,5,...|-0.3122307364002444|
| -11.87816749996684|(10,[0,1,2,3,4,5,...| 0.1338819458914437|
| -11.43180236554046|(10,[0,1,2,3,4,5,...| 0.5248457739492374|
|-11.328415936777782|(10,[0,1,2,3,4,5,...| 0.1542916456260936|
|-11.039347808253828|(10,[0,1,2,3,4,5,...|-1.3518353509995789|
|-10.600130341909033|(10,[0,1,2,3,4,5,...| 0.4030016168294734|
|-10.293714040655924|(10,[0,1,2,3,4,5,...| -1.364529194363915|
| -9.892155927826222|(10,[0,1,2,3,4,5,...| -1.068980736429676|
+-------------------+--------------------+-------------------+
only showing top 20 rows

18/04/17 15:15:56 INFO SparkUI: Stopped Spark web UI at http://192.168.86.21:4040
18/04/17 15:15:56 INFO MapOutputTrackerMasterEndpoint: MapOutputTrackerMasterEndpoint stopped!
18/04/17 15:15:56 INFO MemoryStore: MemoryStore cleared
18/04/17 15:15:56 INFO BlockManager: BlockManager stopped
18/04/17 15:15:56 INFO BlockManagerMaster: BlockManagerMaster stopped
18/04/17 15:15:56 INFO OutputCommitCoordinator$OutputCommitCoordinatorEndpoint: OutputCommitCoordinator stopped!
18/04/17 15:15:56 INFO SparkContext: Successfully stopped SparkContext

---

Run in Docker container

Assuming we have a working local Docker installation execute sbt docker:publishLocal to create the Docker image.

Once the command completes, execute docker images to view the docker image. See output similar to the following:

REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE
sagemaker-spark     0.1                 29dc8b3b2dc8        20 seconds ago      379MB
openjdk             jre-alpine          b1bd879ca9b3        2 months ago        82MB

Now start a container to run the image by executing docker run --rm sagemaker-spark:0.1. You should see output very similar to the output from the local run.

Note: You may see an error related to insuficient memory, like the one shown below. In which case try increasing your docker engine's memory allocation to 4GB.

18/04/05 18:40:29 ERROR SparkContext: Error initializing SparkContext.
java.lang.IllegalArgumentException: System memory 466092032 must be at least 471859200. Please increase heap size using the --driver-memory option or spark.driver.memory in Spark configuration.
      at org.apache.spark.memory.UnifiedMemoryManager$.getMaxMemory(UnifiedMemoryManager.scala:216)
      at org.apache.spark.memory.UnifiedMemoryManager$.apply(UnifiedMemoryManager.scala:198)
      at org.apache.spark.SparkEnv$.create(SparkEnv.scala:330)
      at org.apache.spark.SparkEnv$.createDriverEnv(SparkEnv.scala:174)
      at org.apache.spark.SparkContext.createSparkEnv(SparkContext.scala:257)
      at org.apache.spark.SparkContext.<init>(SparkContext.scala:432)
      at FullModel.Main$.main(Main.scala:17)
      at FullModel.Main.main(Main.scala)

---

Setup Scala Project including Docker

Reference: Lightweight docker containers for Scala apps by Jeroen Rosenberg published August 14, 2017

1. Create an object Main with a main function. Applications that extend scala.App will not work correctly¹.
2. Add the example code.
3. Since the example code does not demonstrate how to establish a SparkContext add the following:

   // Startup
   val conf = new SparkConf()
     .setMaster("local[2]")
     .setAppName("SGD")
     .set("spark.executor.memory", "1g")
   val sc = new SparkContext(conf)

4. Make sure to cleanup the context by adding the following to the example at the end:

   // Shut down
   sc.stop()

5. Update build.sbt to run locally
   1. Scala version must be no greather than 2.11.x²
   2. Spark version should be 2.1.0 and should look like the following:

      libraryDependencies ++= {
        val sparkVer = "2.1.0"
        Seq(
          "org.apache.spark" %% "spark-core" % sparkVer,
          "org.apache.spark" %% "spark-mllib" % sparkVer
        )
      }

   3. Update project/build.properties, set the sbt.version to 0.13.17³
6. Add docker support
   1. Add project/plugins.sbt file with the following contents

      addSbtPlugin("com.typesafe.sbt" % "sbt-native-packager" % "1.2.1")

   2. Update build.sbt
      1. Add the following to the bottom of the file:

         enablePlugins(JavaAppPackaging)
         enablePlugins(DockerPlugin)
         enablePlugins(AshScriptPlugin)
         
         mainClass in Compile := Some("FullModel.Main")
         
         dockerBaseImage := "openjdk:jre-alpine"
         
         mappings in Universal += file("lpsa.data") -> "lpsa.data"

      2. The SBT commands enable plugins that
         1. Allow our app to be packaged as a jar(s) with an executable shell script to run it.
         2. Publish a docker image with the packaged app.
         3. Use ash as the shell instead of bash (needed for alpine based images)
      3. Next we declare the mainClass so that the generated app script knows which class to execute.
      4. Then we instruct the Docker plugin to use a smaller alpine based image rather than the default Debian based image.
      5. Finally we provide a file mapping which instructs the packaging system to include our data file in the staging directory that is later used to construct the image.⁴

---