Difficulties in using a Gcloud Composer DAG to run a Spark job

Question

I'm playing around with Gcloud Composer, trying to create a DAG that creates a DataProc cluster, runs a simple Spark job, then tears down the cluster. I am trying to run the Spark PI example job.

I understand that when calling DataProcSparkOperator I can choose only to define either the main_jar or the main_class property. When I define main_class, the job fails with the error:

java.lang.ClassNotFoundException: org.apache.spark.examples.SparkPi
    at java.net.URLClassLoader.findClass(URLClassLoader.java:382)
    at java.lang.ClassLoader.loadClass(ClassLoader.java:424)
    at java.lang.ClassLoader.loadClass(ClassLoader.java:357)
    at java.lang.Class.forName0(Native Method)
    at java.lang.Class.forName(Class.java:348)
    at org.apache.spark.util.Utils$.classForName(Utils.scala:239)
    at org.apache.spark.deploy.SparkSubmit$.org$apache$spark$deploy$SparkSubmit$$runMain(SparkSubmit.scala:851)
    at org.apache.spark.deploy.SparkSubmit$.doRunMain$1(SparkSubmit.scala:198)
    at org.apache.spark.deploy.SparkSubmit$.submit(SparkSubmit.scala:228)
    at org.apache.spark.deploy.SparkSubmit$.main(SparkSubmit.scala:137)
    at org.apache.spark.deploy.SparkSubmit.main(SparkSubmit.scala)

When I choose to define the main_jar property, the job fails with the error:

Error: No main class set in JAR; please specify one with --class
Run with --help for usage help or --verbose for debug output

I'm at a bit of a loss as to how to resolve this, as I am kinda new to both Spark and DataProc.

My DAG:

import datetime as dt
from airflow import DAG, models
from airflow.contrib.operators import dataproc_operator as dpo
from airflow.utils import trigger_rule

MAIN_JAR = 'file:///usr/lib/spark/examples/jars/spark-examples.jar'
MAIN_CLASS = 'org.apache.spark.examples.SparkPi'
CLUSTER_NAME = 'quickspark-cluster-{{ ds_nodash }}'

yesterday = dt.datetime.combine(
    dt.datetime.today() - dt.timedelta(1),
    dt.datetime.min.time())

default_dag_args = {
    'start_date': yesterday,
    'email_on_failure': False,
    'email_on_retry': False,
    'retries': 1,
    'retry_delay': dt.timedelta(seconds=30),
    'project_id': models.Variable.get('gcp_project')
}

with DAG('dataproc_spark_submit', schedule_interval='0 17 * * *',
    default_args=default_dag_args) as dag:

    create_dataproc_cluster = dpo.DataprocClusterCreateOperator(
        project_id = default_dag_args['project_id'],
        task_id = 'create_dataproc_cluster',
        cluster_name = CLUSTER_NAME,
        num_workers = 2,
        zone = models.Variable.get('gce_zone')
    )

    run_spark_job = dpo.DataProcSparkOperator(
        task_id = 'run_spark_job',
        #main_jar = MAIN_JAR,
        main_class = MAIN_CLASS,
        cluster_name = CLUSTER_NAME
    )

    delete_dataproc_cluster = dpo.DataprocClusterDeleteOperator(
        project_id = default_dag_args['project_id'],
        task_id = 'delete_dataproc_cluster',
        cluster_name = CLUSTER_NAME,
        trigger_rule = trigger_rule.TriggerRule.ALL_DONE
    )

    create_dataproc_cluster >> run_spark_job >> delete_dataproc_cluster

Answer 1

I compared it with a successful job using the CLI and saw that, even when the class was populating the Main class or jar field, the path to the Jar was specified in Jar files:

Checking the operator I noticed there is also a dataproc_spark_jars parameter which is not mutually exclusive to main_class:

run_spark_job = dpo.DataProcSparkOperator(
    task_id = 'run_spark_job',
    dataproc_spark_jars = [MAIN_JAR],
    main_class = MAIN_CLASS,
    cluster_name = CLUSTER_NAME
)

Adding it did the trick: