Distutils Version classes (including LooseVersion) are deprecated

[timokau]

The LooseVersion class (see #579) is now deprecated (distutils commit, setuptools changelog), which results in warnings of the form:

DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.
    if LooseVersion(sklearn.__version__) >= '0.22.0':

We should probably follow the suggestion and move to packaging.version.

[BenjaminBossan]

I didn't know that, thanks for bringing it up.

Is packaging part of the standard Python distribution or an extra dependency for all our supported Python versions? It seems not to exist for < 3.8, could it be? If so, I guess we would need to write the code in a way to still use LooseVersion for old Python versions.

[timokau]

I'm not sure, but as far as I know it's an extra dependency but it should be available for python >= 3.6 according to pypi.

[thomasjpfan]

Using packaging.version is known to be slow: pypa/setuptools#926

For Python >= 3.8, the best way to get versions is: importlib.metadata. Since we still support 3.7, there is a shim on pypi.

If we do not want to add another dependency, we can vendor the code for packaging.version. For example, this is how SciPy vendors it: https://github.com/scipy/scipy/blob/main/scipy/_lib/_pep440.py

[BenjaminBossan]

Why is something that is seemingly simple so complex? :(

For Python >= 3.8, the best way to get versions is: importlib.metadata.

One difference would be that this would compare strings, whereas with LooseVersion, we get a bit of a more sophisticated logic. E.g.:

LooseVersion(sklearn.__version__).version # => [1, 2, 'dev', 0]

Maybe it makes no difference for what we use, but I would much prefer a solution that uses the existing logic to compare versions.

For example, this is how SciPy vendors it: https://github.com/scipy/scipy/blob/main/scip

The link is broken but in general, if we're just dealing with a few lines of code, I'd prefer vendoring to adding a new dependency.

[thomasjpfan]

My link got cut out. Here is the full link: https://github.com/scipy/scipy/blob/main/scipy/_lib/_pep440.py

[BenjaminBossan]

The scipy code looks good and should be battle tested. I'm not to happy that there seems to be no better option, but vendoring that code gets a +1 from me.

[BenjaminBossan]

With the module copied from scipy, we should be good, so let's close for now.

[timokau]

Thank you!