Spark 2.1 with cluster mode ignores extraJavaOptions

Question

We are running a HortonWorks Spark Distribution 2.1. We noticed that when we run Spark with yarn-cluster as master, whatever we pass as executors.extraJavaOptions is completely ignored, and the job maintains the default values in the executors (as seen through the Spark UI).

I was wondering whether anyone has noticed anything similar and has some wisdom to share. I would be glad to provide any more details needed.

Update: Here is the command executed, along with the options, as requested:

spark-submit --name streaming-driver --num-executors 40 
    --conf spark.metrics.conf=streaming.metrics.properties 
    --files kafka_client_jaas.conf#kafka_client_jaas.conf,
            appconf.conf#appconf.conf,
            streaming.metrics.properties#streaming.metrics.properties,
            kafka.conf#kafka.conf,log4j.properties#log4j.properties 
    --conf "spark.executor.extraJavaOptions=
               -Dcom.sun.management.jmxremote 
               -Dcom.sun.management.jmxremote.port=0 
               -Dcom.sun.management.jmxremote.rmi.port=0 
               -Dcom.sun.management.jmxremote.authenticate=false 
               -Dcom.sun.management.jmxremote.ssl=false 
               -Dcom.sun.management.jmxremote.local.only=false 
               -Djava.security.auth.login.config=./kafka_client_jaas.conf 
               -XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode" 
    --conf spark.metrics.conf=streaming.metrics.properties 
    --conf spark.hadoop.yarn.timeline-service.enabled=false 
    --driver-java-options=" 
           -Dhdp.version=2.5.3.0-37 -Dconfig.file=appconf.conf 
           -XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode 
           -Dlog4j.configuration=log4j.properties 
           -Djava.security.auth.login.config=./kafka_client_jaas.conf 
           -Dcom.sun.management.jmxremote 
           -Dcom.sun.management.jmxremote.port=0 
           -Dcom.sun.management.jmxremote.rmi.port=0 
           -Dcom.sun.management.jmxremote.authenticate=false 
           -Dcom.sun.management.jmxremote.ssl=false 
           -Dcom.sun.management.jmxremote.local.only=false" 
    --master yarn-cluster 
    --driver-memory "6g" 
    --properties-file spark.conf --class < class> application.jar

Note this is slightly modified because of some other security parameters like keytabs etc. No newline exists as well, put here only so that it is more readable.

How do you specify the options? Edit your question with the command line. — Jacek Laskowski, Dec 03 '17 at 13:33

score 0 · Answer 1 · answered Dec 03 '17 at 17:39

Looks like you are wrapping spark.executor.extraJavaOptions proery also inside double inverted quotes. I think only the value should be inside quotes some thing like this:

 --conf spark.executor.extraJavaOptions="
                   -Dcom.sun.management.jmxremote 
                   -Dcom.sun.management.jmxremote.port=0 
                   -Dcom.sun.management.jmxremote.rmi.port=0 
                   -Dcom.sun.management.jmxremote.authenticate=false 
                   -Dcom.sun.management.jmxremote.ssl=false 
                   -Dcom.sun.management.jmxremote.local.only=false 
                   -Djava.security.auth.login.config=./kafka_client_jaas.conf 
                   -XX:+UseConcMarkSweepGC -XX:+CMSIncrementalMode"

Unfortunately, I have tried to do as you suggested as well, still the same error. — Y.K., Dec 03 '17 at 19:31

Spark 2.1 with cluster mode ignores extraJavaOptions

1 Answers1