cluster.md

Table of Contents generated with DocToc

• Configuring Job Server for YARN cluster mode
  • Job Server configuration
  • Reading files uploaded via frontend

Configuring Job Server for YARN cluster mode

See also running YARN in client mode, running YARN on EMR and running on Mesos.

Job Server configuration

Add the following properties in your job server config file:

• set spark.master property to yarn, spark://... or mesos://...
• set spark.submit.deployMode property to cluster
• set spark.jobserver.context-per-jvm to true
• set akka.remote.netty.tcp.hostname to the cluster interface of the host running the frontend
• set akka.remote.netty.tcp.maximum-frame-size to support big remote jars fetch
• set spark.jobserver.network-address-resolver to a suitable value according to your environment. akka is recommended.

Optional / required in spark standalone mode:

• set REMOTE_JOBSERVER_DIR to hdfs://..., file://... or http://... in your settings xxx.sh
• copy spark-job-server.jar, your job server config and log4j-cluster.properties file into this location

Example job server config (replace CLUSTER-IP with the internal IP of the host running the job server frontend):

spark {
  # deploy in yarn cluster mode
  master = yarn
  submit.deployMode = cluster

  jobserver {
    context-per-jvm = true

    # start a H2 DB server, reachable in your cluster
    sqldao {
      jdbc {
        url = "jdbc:h2:tcp://CLUSTER-IP:9092/h2-db;AUTO_RECONNECT=TRUE"
      }
    }
    startH2Server = false
  }
}

# start akka on this interface, reachable from your cluster
akka {
  remote.netty.tcp {
    hostname = "CLUSTER-IP"

    # This controls the maximum message size, including job results, that can be sent
    maximum-frame-size = 100 MiB
  }
}

Note:

• YARN transfers the files provided via --files submit option into the cluster / container. Spark standalone does not support this in cluster mode and you have to transfer them manual.
• Instead of running a H2 DB instance you can also run a real DB reachable inside your cluster. You can't use the default (host only) H2 configuration in a cluster setup.
• Akka binds by default to the local host interface and is not reachable from the cluster. You need to configure the akka hostname to the cluster internal address.
• At least one slave node needs to be attached to the master for contexts to be successfully created in Cluster Mode.

Reading files uploaded via frontend

Files uploaded via the data API (/data) are stored on your job server frontend host. Call the DataFileCache API implemented by the job environment in your spark jobs to access them:

  object RemoteDriverExample extends NewSparkJob {
    def runJob(sc: SparkContext, runtime: JobEnvironment, data: JobData): JobOutput =
      runtime.getDataFile(...)

The job server transfers the files via akka to the host running your driver and caches them there.

Note: Files uploaded via the JAR or binary API are stored and transfered via the Job DB.

Configuring Job Server for supervised cluster mode

• See the supervise-mode document