Launching Spark on YARN

Experimental support for running over a YARN (Hadoop NextGen) cluster was added to Spark in version 0.6.0. Because YARN depends on version 2.0 of the Hadoop libraries, this currently requires checking out a separate branch of Spark, called yarn, which you can do as follows:

git clone git://github.com/mesos/spark
cd spark
git checkout -b yarn --track origin/yarn

Preparations

In order to distribute Spark within the cluster, it must be packaged into a single JAR file. This can be done by running sbt/sbt assembly
Your application code must be packaged into a separate JAR file.

If you want to test out the YARN deployment mode, you can use the current Spark examples. A spark-examples_2.9.3-0.7.3 file can be generated by running sbt/sbt package. NOTE: since the documentation you’re reading is for Spark version 0.7.3, we are assuming here that you have downloaded Spark 0.7.3 or checked it out of source control. If you are using a different version of Spark, the version numbers in the jar generated by the sbt package command will obviously be different.

Launching Spark on YARN

The command to launch the YARN Client is as follows:

SPARK_JAR=<SPARK_YAR_FILE> ./run spark.deploy.yarn.Client \
  --jar <YOUR_APP_JAR_FILE> \
  --class <APP_MAIN_CLASS> \
  --args <APP_MAIN_ARGUMENTS> \
  --num-workers <NUMBER_OF_WORKER_MACHINES> \
  --worker-memory <MEMORY_PER_WORKER> \
  --worker-cores <CORES_PER_WORKER>

For example:

SPARK_JAR=./core/target/spark-core-assembly-0.7.3.jar ./run spark.deploy.yarn.Client \
  --jar examples/target/scala-2.9.3/spark-examples_2.9.3-0.7.3.jar \
  --class spark.examples.SparkPi \
  --args standalone \
  --num-workers 3 \
  --worker-memory 2g \
  --worker-cores 2

The above starts a YARN Client programs which periodically polls the Application Master for status updates and displays them in the console. The client will exit once your application has finished running.

Important Notes

When your application instantiates a Spark context it must use a special “standalone” master url. This starts the scheduler without forcing it to connect to a cluster. A good way to handle this is to pass “standalone” as an argument to your program, as shown in the example above.
YARN does not support requesting container resources based on the number of cores. Thus the numbers of cores given via command line arguments cannot be guaranteed.