on Amazon EMR 4.0.0, setting /etc/spark/conf/spark-env.conf is ineffective

Question

I'm launching my spark-based hiveserver2 on Amazon EMR, which has an extra classpath dependency. Due to this bug in Amazon EMR:

https://petz2000.wordpress.com/2015/08/18/get-blas-working-with-spark-on-amazon-emr/

My classpath cannot be submitted through "--driver-class-path" option

So I'm bounded to modify /etc/spark/conf/spark-env.conf to add the extra classpath:

# Add Hadoop libraries to Spark classpath
SPARK_CLASSPATH="${SPARK_CLASSPATH}:${HADOOP_HOME}/*:${HADOOP_HOME}/../hadoop-hdfs/*:${HADOOP_HOME}/../hadoop-mapreduce/*:${HADOOP_HOME}/../hadoop-yarn/*:/home/hadoop/git/datapassport/*"

where "/home/hadoop/git/datapassport/*" is my classpath.

However after launching the server successfully, the Spark environment parameter shows that my change is ineffective:

spark.driver.extraClassPath :/usr/lib/hadoop/*:/usr/lib/hadoop/../hadoop-hdfs/*:/usr/lib/hadoop/../hadoop-mapreduce/*:/usr/lib/hadoop/../hadoop-yarn/*:/etc/hive/conf:/usr/lib/hadoop/../hadoop-lzo/lib/*:/usr/share/aws/emr/emrfs/conf:/usr/share/aws/emr/emrfs/lib/*:/usr/share/aws/emr/emrfs/auxlib/*

Is this configuration file obsolete? Where is the new file and how to fix this problem?

Answer 1

You can use the --driver-classpath.

Start a spark-shell on the master node from a fresh EMR cluster.

spark-shell --master yarn-client
scala> sc.getConf.get("spark.driver.extraClassPath")
res0: String = /etc/hadoop/conf:/usr/lib/hadoop/*:/usr/lib/hadoop-hdfs/*:/usr/lib/hadoop-yarn/*:/usr/lib/hadoop-lzo/lib/*:/usr/share/aws/aws-java-sdk/*:/usr/share/aws/emr/emrfs/conf:/usr/share/aws/emr/emrfs/lib/*:/usr/share/aws/emr/emrfs/auxlib/*

Add your JAR files to the EMR cluster using a --bootstrap-action.

When you call spark-submit prepend (or append) your JAR files to the value of extraClassPath you got from spark-shell

spark-submit --master yarn-cluster --driver-classpath /home/hadoop/my-custom-jar.jar:/etc/hadoop/conf:/usr/lib/hadoop/*:/usr/lib/hadoop-hdfs/*:/usr/lib/hadoop-yarn/*:/usr/lib/hadoop-lzo/lib/*:/usr/share/aws/aws-java-sdk/*:/usr/share/aws/emr/emrfs/conf:/usr/share/aws/emr/emrfs/lib/*:/usr/share/aws/emr/emrfs/auxlib/*

This worked for me using EMR release builds 4.1 and 4.2.

The process for building spark.driver.extraClassPath may change between releases, which may be the reason why SPARK_CLASSPATH doesn't work anymore.

Answer 2

Have you tried setting spark.driver.extraClassPath in spark-defaults? Something like this:

[
  {
    "Classification": "spark-defaults",
    "Properties": {
      "spark.driver.extraClassPath": "${SPARK_CLASSPATH}:${HADOOP_HOME}/*:${HADOOP_HOME}/../hadoop-hdfs/*:${HADOOP_HOME}/../hadoop-mapreduce/*:${HADOOP_HOME}/../hadoop-yarn/*:/home/hadoop/git/datapassport/*"
    }
  }
]