Remove use of entrypoints

[s-t-e-v-e-n-k]

Since Python 3.8, the standard library has included functionality to query entry points directly using importlib.metadata. Since the API has changed for the better with Python 3.10, we need to support both ways of using it.

[joshmoore]

Thanks, @s-t-e-v-e-n-k. I've triggered the workflows.

[codecov]

Codecov Report

Merging #442 (af20af8) into main (d667b31) will not change coverage.
The diff coverage is 100.00%.

@@            Coverage Diff            @@
##              main      #442   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           56        56           
  Lines         2243      2242    -1     
=========================================
- Hits          2243      2242    -1     

Files Changed	Coverage Δ
numcodecs/tests/test_entrypoints.py	100.00% <ø> (ø)
numcodecs/registry.py	100.00% <100.00%> (ø)

[joshmoore]

Makes sense to me. I've also opened (a currently failing) PR to remove entrypoint from the conda-forge package: conda-forge/numcodecs-feedstock#95

[jakirkham]

cc @martindurant

[martindurant]

I don't know anything about `.select, but I trust you. We actually have tests exercising the entrypoint mechanism, right?

[s-t-e-v-e-n-k]

There are tests, but they are marked as xfail.

[joshmoore]

Good catch, @s-t-e-v-e-n-k. Can you un-xfail them here and we'll address whatever issues there are in the wheels build?

[s-t-e-v-e-n-k]

I have rebased and removed the xfail.

[joshmoore]

Re-launched workflows.

[jakirkham]

Looks like the entrypoint test is failing with wheels

[joshmoore]

      @pytest.fixture()
      def set_path():
          sys.path.append(here)
          numcodecs.registry.run_entrypoints()
          yield
          sys.path.remove(here)
          numcodecs.registry.run_entrypoints()
  >       numcodecs.registry.codec_registry.pop("test")
  E       KeyError: 'test'

This may be similar to the setup needed for conda-forge testing: i.e. it's not enough to just sys.path, but the test must actually be installed due to differing runtime environments.

[s-t-e-v-e-n-k]

This may be similar to the setup needed for conda-forge testing: i.e. it's not enough to just sys.path, but the test must actually be installed due to differing runtime environments.

I've been wondering about that, if we want to test this piece meal by mocking, and also doing an integration test by doing ... something like what the sys.path manipulation is trying to accomplish.

[martindurant]

You should need both the package and a distinfo on the path, like intake does: https://github.com/intake/intake/tree/master/intake/source/tests/plugin_searchpath

[s-t-e-v-e-n-k]

I fell down an importlib.metadata vs entrypoints rabbit hole, but I nailed it down to the wheel not including the needed files for the tests to pass. I'm not certain why MANIFEST.in is not helping us here.

[jakirkham]

RTD failures are due to an upstream issue ( readthedocs/readthedocs.org#10601 ). Mentioned these instances there

[jakirkham]

Looks like it may have been a GH issue. Restarted RTD and it appears to have passed. Also enabled GHA jobs to run

[joshmoore]

Hurrah! 🎉

[joshmoore]

Thanks for all the hard work, @s-t-e-v-e-n-k! Merging lest other RTD issues crop up, but if there are any concerns from anyone, just let us know before the upcoming release.