Drop Python 3.9

[Armavica]

Description

Related Issue

• Closes Drop support for Python 3.9 #680
• Related to #

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

[maresb]

Looks fine, but I'm wondering why? Is 3.9 that old??? 👴

[Armavica]

Looks fine, but I'm wondering why? Is 3.9 that old??? 👴

This might be a reason: https://scientific-python.org/specs/spec-0000/

[ricardoV94]

Looks fine, but I'm wondering why? Is 3.9 that old??? 👴

This might be a reason: https://scientific-python.org/specs/spec-0000/

Yes that's the reason

[maresb]

Makes sense, though sometimes it feels like the world moves too quickly.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 76.28571% with 83 lines in your changes are missing coverage. Please review.

Project coverage is 80.86%. Comparing base (60e2510) to head (95ac1c6).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #682      +/-   ##
==========================================
+ Coverage   80.83%   80.86%   +0.02%     
==========================================
  Files         162      162              
  Lines       46830    46793      -37     
  Branches    11447    11442       -5     
==========================================
- Hits        37857    37838      -19     
+ Misses       6710     6696      -14     
+ Partials     2263     2259       -4     

Files	Coverage Δ
pytensor/compile/compilelock.py	97.05% <ø> (-0.09%)	⬇️
pytensor/compile/function/pfunc.py	82.92% <100.00%> (ø)
pytensor/compile/mode.py	84.40% <100.00%> (ø)
pytensor/compile/monitormode.py	63.04% <ø> (ø)
pytensor/compile/profiling.py	76.32% <100.00%> (ø)
pytensor/compile/sharedvalue.py	93.75% <100.00%> (ø)
pytensor/configparser.py	86.36% <100.00%> (-0.06%)	⬇️
pytensor/graph/destroyhandler.py	69.20% <100.00%> (ø)
pytensor/graph/op.py	87.50% <100.00%> (ø)
pytensor/graph/replace.py	84.21% <100.00%> (ø)
... and 77 more

... and 2 files with indirect coverage changes

[maresb]

I knew there was some reason I had a bad feeling about this. I marked this as draft as a safety measure so that nobody merges before we discuss a little bit.

There's currently a bottleneck because on conda-forge we rebuilt PyTensor with support for Python 3.12 only since PyTensor 2.19. But we're pinning pytensor <2.19 in PyMC. So if we were to merge this now, then we'd only support Python 3.10 and 3.11, which feels a bit narrow to me.

I propose that we first cut a release that supports PyTensor 2.19. Then we'll have a version of PyMC that works for the full range 3.9-3.12. (Maybe it doesn't matter, but my intuition is that having a release with a nice range like this makes it a lot easier to satisfy dependencies.) Then we can immediately follow that up with another release dropping 3.9 support.

[ricardoV94]

I don't care, but the whole point of that spec is that no scientific package has to support more than 3 major python versions (assuming one per year I guess).

And we were already restricted to 3 up until pretty much now

[maresb]

Sorry, I butchered the versions in my above now-edited comment (or I'm not thinking straight). If we merge this right now, we'll only be supporting 2 major Python versions.

With my suggestion, we'll have one version that supports 4 and then drop down to a nominal 3 as you state.

[ricardoV94]

This will be incorporated as a "major release", so PyMC will only be able to depend on it when we manually bump the dependency limit over there, so it can't have any effect?

[maresb]

Ah... ok, I think that makes sense.

You're saying that we merge this and release it as PyTensor 2.20. Then over in PyMC, completely independently we increase the range to include PyTensor 2.19 and support Python 3.9-3.12 which matches my suggestion. Then, again in PyMC completely independently when we pin to PyTensor 2.20, we support the nominal Python 3.10-3.12 as per scientific python spec.

Thanks for walking me through that. 😅

EDIT: Fixed the logic to be in line with 2.19/2.20 as pointed out in the next comment

[ricardoV94]

Yes except we are not using semver so major changes happen on intermediate numbers. The logic you mentioned would be with 2.19/2.20

[ricardoV94]

PyPI / Build wheels on ubuntu-20.04 failed, I restarted it.

[ricardoV94]

Thanks @Armavica!