(Apache Beam) Cannot increase executor memory - it is fixed at 1024M despite using multiple settings

Question

I am running an apache beam workload on Spark. I initialized the workers with 32GB of memory (slave run with -c 2 -m 32G). Spark submit sets driver memory to 30g and executor memory to 16g. However, executors fail with java.lang.OutOfMemoryError: Java heap space.

The master gui indicates that memory per executor is 1024M. In addition, I see that all java processes are launched with -Xmx 1024m. This means spark-submit doesn't propagate it's executor settings to the executors.

Pipeline options are as follows:

   --runner PortableRunner \
    --job_endpoint=localhost:8099 \
    --environment_type=PROCESS \
    --environment_config='{"command": "$HOME/beam/sdks/python/container/build/target/launcher/linux_amd64/boot"}'

Job endpoint is setup in the default way: docker run --rm --network=host --name spark-jobservice apache/beam_spark_job_server:latest --spark-master-url=spark://$HOSTNAME:7077

How do I make sure the settings propagate to the executors?

Update: I set conf/spark-defaults.conf to

spark.driver.memory 32g
spark.executor.memory 32g

and conf/spark-env.sh to

SPARK_EXECUTOR_MEMORY=32g

and restarted the cluster and relaunched everything, and executor memory is still limited to 1024M

Nature

same exact problem here, the Beam pipeline doesn't seem to take into account spark defaults for executor memory. – Dibble

Answer 1

I found the reason and a workaround.

The jobserver container is running internally its own spark distribution, so the settings configured in the spark distribution on your local machine have no effect.

The solution is thus to change the configuration in the jobserver container, for instance by passing the corresponding environment variable when launching it:

docker run -e SPARK_EXECUTOR_MEMORY=32g --rm --network=host --name spark-jobservice apache/beam_spark_job_server:latest --spark-master-url=spark://$HOSTNAME:7077

Dibble