[jira] [Created] (SPARK-28484) spark-submit uses wrong SPARK_HOME with deploy-mode "cluster"

["Kalle Jepsen (JIRA)" <ji...@apache.org>]

Kalle Jepsen created SPARK-28484:
------------------------------------

             Summary: spark-submit uses wrong SPARK_HOME with deploy-mode "cluster"
                 Key: SPARK-28484
                 URL: https://issues.apache.org/jira/browse/SPARK-28484
             Project: Spark
          Issue Type: Bug
          Components: Deploy
    Affects Versions: 2.4.3
            Reporter: Kalle Jepsen


When submitting an application jar to a remote Spark cluster with spark-submit and deploy-mode = "cluster", the driver command that is issued on one of the workers seems to be configured with the SPARK_HOME of the local machine, from which spark-submit was called, not the one where the driver is actually running.

 

I.e. if I have spark installed locally under e.g. /opt/apache-spark and hadoop under /usr/lib/hadoop-3.2.0, but the cluster administrator installs spark under /usr/local/spark on the workers, the command that is issued on the worker still looks sth like this:

 

{{"/usr/lib/jvm/java/bin/java" "-cp" "/opt/apache-spark/conf:/etc/hadoop:/usr/lib/hadoop-3.2.0/share/hadoop/common/lib/*:/usr/lib/hadoop-3.2.0/share/hadoop/common/*:/usr/lib/hadoop-3.2.0/share/hadoop/hdfs:/usr/lib/hadoop-3.2.0/share/hadoop/hdfs/lib/*:/usr/lib/hadoop-3.2.0/share/hadoop/hdfs/*:/usr/lib/hadoop-3.2.0/share/hadoop/mapreduce/lib/*:/usr/lib/hadoop-3.2.0/share/hadoop/mapreduce/*:/usr/lib/hadoop-3.2.0/share/hadoop/yarn:/usr/lib/hadoop-3.2.0/share/hadoop/yarn/lib/*:/usr/lib/hadoop-3.2.0/share/hadoop/yarn/*" "-Xmx1024M" "-Dspark.jars=file:///some/application.jar" "-Dspark.driver.supervise=false" "-Dspark.submit.deployMode=cluster" "-Dspark.master=spark://<SPARK_MASTER>:7077" "-Dspark.app.name=<APPNAME>" "-Dspark.rpc.askTimeout=10s" "org.apache.spark.deploy.worker.DriverWrapper" "spark://Worker@<WORKER_HOST>:65000" "/some/application.jar" "some.class.Name"}}

 

Is this expected behavior and/or can I somehow control that?

 

Steps to reproduce:

 

1. Install Spark locally (with a SPARK_HOME that's different on the cluster)

{{2. Run: spark-submit --deploy-mode "cluster" --master "spark://spark.example.com:7077" --class "com.example.SparkApp" "hdfs:/some/application.jar"}}

3. Observe that the application fails because some spark and/or hadoop classes cannot be found

 

This applies to Spark Standalone, I haven't tried with YARN



--
This message was sent by Atlassian JIRA
(v7.6.14#76016)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org