Cannot run Python PortableRunner -> SparkRunner on EMR cluster

[damccorm]

I have been trying to run the python word-count example on an AWS EMR cluster. And it does not work.

Things I have tried:

• Running with 

python3 py_codes/word_count_beam.py --output word_count_output --runner=SparkRunner

This results in implicitly running with --spark-master-url local[4] which defeats the purpose of running it in a cluster

• Tried

python3 py_codes/word_count_beam.py --output word_count_output --runner=SparkRunner --spark-master-url=yarn

Still uses local master.

• Could not use method described in https://beam.apache.org/documentation/runners/spark/ under "Running on a pre-deployed Spark cluster" because in yarn master is not exposed with an URL like localhost:7077

• Tried

python3 py_codes/word_ount_beam.py --output word_count_output --runner=SparkRunner --output_executable_path=jars/beam_word_count.jar

as described in https://issues.apache.org/jira/browse/BEAM-8970
It can create a jar file, but when I submit the jar with spark-submit I get docker permission denied exception. Possibly related to https://issues.apache.org/jira/browse/BEAM-6020

So, no way to run a python beam code in a yarn spark cluster?
This also means no way to run TFX code (which uses beam) in a yarn cluster.

Imported from Jira BEAM-11378. Original Jira may contain additional context.
Reported by: ratulray.

[moradology]

It appears to me that the PROCESS SDK environment avoids the problems encountered here. A discussion of what ended up working for me on this issue from a downstream project: pangeo-forge/pangeo-forge-runner#133