Resolution: Not A Problem
The goal of this ticket is to ship and use custom code inside the spark executors using PEX
This currently works fine with conda (disadvantages are that you have a separate conda package repo and ship the python interpreter all the time)
Basically the workflow is
- to zip the local conda environment (conda pack also works)
- ship it to each executor as an archive
- modify PYSPARK_PYTHON to the local conda environment
I think it can work the same way with virtual env. There is the SPARK-13587 ticket to provide nice entry points to spark-submit and SparkContext but zipping your local virtual env and then just changing PYSPARK_PYTHON env variable should already work.
I also have seen this blogpost. But recreating the virtual env each time doesn't seem to be a very scalable solution. If you have hundreds of executors it will retrieve the packages on each excecutor and recreate your virtual environment each time. Same problem with this proposal
SPARK-16367 from what I understood.
Another problem with virtual env is that your local environment is not easily shippable to another machine. In particular there is the relocatable option (see https://virtualenv.pypa.io/en/stable/userguide/#making-environments-relocatable, https://stackoverflow.com/questions/7153113/virtualenv-relocatable-does-it-really-work) which makes it very complicated for the user to ship the virtual env and be sure it works.
And here is where pex comes in. It is a nice way to create a single executable zip file with all dependencies included. You have the pex command line tool to build your package and when it is built you are sure it works. This is in my opinion the most elegant way to ship python code (better than virtual env and conda)
The problem why it doesn't work out of the box is that there can be only one single entry point. So just shipping the pex files and setting PYSPARK_PYTHON to the pex files doesn't work. You can nevertheless tune the env variable PEX_MODULE and runtime to provide different entry points.