Test against mindeps

[ivirshup]

• Closes Minimum dependencies test job #1203
• Tests added
• Release note added (or unnecessary)

TODO:

• Get all tests running

Won't do in this PR:

• GPU support (it's going to be more difficult with the self-hosted runner + gotta get cuda versions working right
• Fancier minimum version detection

Current problems:

• Warnings are different in earlier versions...
• Pandas has different results for pd.value_counts depending on version
• Should this be collecting coverage to check on version bounded things?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (b9f346c) 85.83% compared to head (9b8816b) 85.87%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1314      +/-   ##
==========================================
+ Coverage   85.83%   85.87%   +0.03%     
==========================================
  Files          35       35              
  Lines        5584     5584              
==========================================
+ Hits         4793     4795       +2     
+ Misses        791      789       -2     

Flag	Coverage Δ
gpu-tests	52.38% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files	Coverage Δ
anndata/experimental/merge.py	91.28% <ø> (ø)

... and 1 file with indirect coverage changes

[ivirshup]

This last doctest is a pain (concat_on_disk), because pandas actually changed the output for value_counts between v1 and v2.

I'd rather not just skip it, but unsure if there's a good solution here. Ideally we'd just ignore the lines that are different or skip this bit only when pandas is older.

We could maybe skip doc tests for minimum dependencies since I would expect more things like this to come up (especially with warnings), but they are very good for catching bugs.

[flying-sheep]

Suggested change

	if shape[axis] == 0 and Version(pd.__version__) >= Version("2.1")
	if shape[axis] == 0 and Version(version("pandas")) >= Version("2.1")

same as above

[flying-sheep]

This last doctest is a pain (concat_on_disk), because pandas actually changed the output for value_counts between v1 and v2. I'd rather not just skip it, but unsure if there's a good solution here.

What’s the difference in the format? maybe we can modify it in a way that smooths over the differences, e.g. dict(sorted(x.items())) if the order is different or something.

[ivirshup]

The name of the series and the name of the index are different. For example:

>>> pd.DataFrame({"a": [1, 1, 2, 3, 4, 4, 4]})["a"].value_counts()
4    3
1    2
2    1
3    1
Name: a, dtype: int64
>>> pd.__version__
'1.4.4'

>>> pd.DataFrame({"a": [1, 1, 2, 3, 4, 4, 4]})["a"].value_counts()
a
4    3
1    2
2    1
3    1
Name: count, dtype: int64
>>> pd.__version__
'2.1.2'

I don't like the idea of modifying the command just to get it reproducible since it's meant to be an example for the user, and it's a weird thing to select as an example without the context.

If we had a different example to show, that could be useful. We could even just remove this line.

[flying-sheep]

I don't like the idea of modifying the command just to get it reproducible since it's meant to be an example for the user, and it's a weird thing to select as an example without the context.

Sure, only if it makes sense. a trailing .to_dict() might!

[ivirshup]

I'm converting back to draft while I get this working for scanpy. I think there will be a few changes that it'd be good to get in here too.

[flying-sheep]

OK! Thanks for the hard work figuring out the deps.

I guess we should make the script a standalone package or something if you want to reuse it?