I have an archive (basically a bundled conda environment + my application) which I can easily use with pyspark in yarn master mode:
PYSPARK_PYTHON=./pkg/venv/bin/python3 \
spark-submit \
--conf spark.yarn.appMasterEnv.PYSPARK_PYTHON=./pkg/venv/bin/python3 \
--master yarn \
--deploy-mode cluster \
--archives hdfs:///package.tgz#pkg \
app/MyScript.py
This works as expected, no surprise here.
Now how could I run this if MyScript.py is inside package.tgz. not on my local filesystem?
I would like to replace the last line of my command with eg. ./pkg/app/MyScript.py
but then spark complains: java.io.FileNotFoundException: File file:/home/blah/pkg/app/MyScript.py does not exist
.
I could of course extract it first, put it separately on hdfs... There are workarounds but as I have everything in one nice place, I would love to use it.
If it's relevant, this is spark 2.4.0, python 3.7, on CDH.
As I understand it, you cannot: you must supply a Python script to spark-submit
.
But you can have a very short script and use --py-files
to distribute a ZIP or EGG of the rest of your code:
# go.py
from my.app import run
run()
# my/app.py
def run():
print("hello")
You can create a ZIP file containing the my directory and submit that with the short entry point script: spark-submit --py-files my.zip go.py
If you like, you can make a generic go.py
that accepts arguments telling it which module and method to import and run.