Run a Spark Job

Running a Spark job

  1. Before submitting your job, upload the artifact (such as a jar file) to a location that is visible to the cluster (such as HTTP, S3, or HDFS). Learn more.

  2. Run the job.

    • Include all configuration flags for the job before the jar url.
    • Provide the arguments for the Spark job after the jar url.

    Follow the template dcos spark run --submit-args="<flags> URL [args], where:

    • <flags> are options like --conf spark.cores.max=16 and --class my.aprk.App
    • URL is the location of the application -[args] are any arguments for the application

    For example:

     dcos spark run --submit-args=--class MySampleClass"
     dcos spark run --submit-args="--py-files"
     dcos spark run --submit-args=""

    If your job runs successfully, you will get a message with the job’s submission ID:

     Run job succeeded. Submission id: driver-20160126183319-0001
  3. View the Spark scheduler progress by navigating to the Spark dispatcher at http:///service/spark/`.

  4. View the job’s logs through the Mesos UI at http://<dcos-url>/mesos/ or by running dcos task log --follow <submission_id>.

Setting Spark properties

Spark job settings are controlled by configuring Spark properties.


All properties are submitted through the --submit-args option to dcos spark run. There are a few options that are unique to DC/OS that are not in Spark Submit (for example --keytab-secret-path). View dcos spark run --help for a list of all these options. All --conf properties supported by Spark can be passed through the command-line with within the --submit-args string.

dcos spark run --submit-args="--conf spark.executor.memory=4g --supervise --class MySampleClass 30"

Setting automatic configuration defaults

To set Spark properties with a configuration file, create a spark-defaults.conf file and set the environment variable SPARK_CONF_DIR to the containing directory. Learn more.

Using a properties file

To reuse Spark properties without cluttering the command line, the CLI supports passing a path to a local file containing Spark properties. Such a file consists of properties and values separated by whitespace. For example:

spark.mesos.containerizer   mesos
spark.executors.cores       4
spark.eventLog.enabled      true
spark.eventLog.dir          hdfs:///history

This sample property file sets the containerizer to mesos, the executor cores to 4 and enables the history server. This file is parsed locally, so it is not be available to your driver applications.


Enterprise DC/OS provides a secrets store to enable access to sensitive data such as database passwords, private keys, and API tokens. DC/OS manages secure transportation of secret data, access control and authorization, and secure storage of secret content. A secret can be exposed to drivers and executors as a file or as an environment variable.

To configure a job to access a secret, see the sections on

DC/OS overlay network

To submit a Spark job inside the DC/OS Overlay Network, run a command similar to the following:

dcos spark run --submit-args="--conf spark.mesos.containerizer=mesos --conf --class MySampleClass"

Note that DC/OS overlay support requires the UCR rather than the default Docker Containerizer, so you must set --conf spark.mesos.containerizer=mesos.

Driver failover timeout

The --conf spark.mesos.driver.failoverTimeout option specifies the number of seconds that the master will wait for the driver to reconnect, after being temporarily disconnected, before it tears down the driver framework by killing all its executors. The default value is zero, meaning no timeout. If the driver disconnects and the value of this option is zero, the master immediately tears down the framework.

To submit a job with a nonzero failover timeout, run a command similar to the following:

dcos spark run --submit-args="--conf spark.mesos.driver.failoverTimeout=60 --class MySampleClass"

IMPORTANT: If you kill a job before it finishes, the framework will persist as an inactive framework in Mesos for a period equal to the failover timeout. You can manually tear down the framework before that period is over by hitting the Mesos teardown endpoint.


The DC/OS Apache Spark Docker image contains OpenJDK 8 and Python 2.7.6.

DC/OS Apache Spark distributions 1.X are compiled with Scala 2.10. DC/OS Apache Spark distributions 2.X are compiled with Scala 2.11. Scala is not binary compatible across minor verions, so your Spark job must be compiled with the same Scala version as your version of DC/OS Apache Spark.

The default DC/OS Apache Spark distribution is compiled against Hadoop 2.9 libraries. However, you can choose a different version by following the instructions in Customize Spark distribution.