[RLlib][Tune] Build Ray-ML docker image for Python 3.9

[mwtian]

What is the problem?

Currently building Ray docker image with Python 3.9 times out, because some of the direct and transitive dependencies pinned in python/requirements/rllib/requirements_rllib.txt and python/requirements/tune/requirements_tune.txt do not support Python 3.9, and the pip dependency resolver goes into an edge case.

Ray version and other system information (Python version, TensorFlow version, OS):

Python 3.9, Ray 1.4.1

Reproduction (REQUIRED)

https://github.com/ray-project/ray/pull/16571/files

[amogkam]

Hmm can we just not build Ray-ML Docker images for 3.9 for now?

The standard Ray docker images shouldn’t need requirements_rllib and requirements_tune

[amogkam]

Also what’s the exact error you are getting?

another alternative is to just skip over the libraries that don’t have 3.9 support- something like this https://stackoverflow.com/questions/22250483/stop-pip-from-failing-on-single-package-when-installing-with-requirements-txt

[mwtian]

If it is ok with others to not build Ray-ML Docker images for 3.9, that can definitely be a solution.
The error I'm currently getting, after trying to update some dependencies, is https://buildkite.com/ray-project/ray-builders-pr/builds/7597#9190b10f-ce4d-428f-8c5e-c1e9d3835813/6-2858

[krfricke]

cc @richardliaw do we have a target on when to fully support Python 3.9?