Use unmangled extra in DistInfoDistribution._compute_dependencies

[blueyed]

With an extra of "GFK" this would not be found as "gfk".

I came up with this when investigating an issue with pip-tools / pip (pypa/pip#3661).

[jaraco]

Hi @blueyed, thanks for the suggestion. I'm not sure I fully understand the implications of this change. I hope you can help.

So your use case suggests that something doesn't work in some circumstances, but as I walk through the tickets, I don't see a use case that I can readily use to replicate the failure (preferably through setuptools directly, without pip). Could you put something like that together to demonstrate the undesirable behavior?

The use of safe_extra seems relevant here. By lower-casing it, it's normalizing it, which by some measures would mean that the extras are compared case-insensitive. Perhaps that was the intention of the original code. Your change makes it specifically case sensitive. Another way to think about the issue is that marker.evaluate should compare markers case-insensitive. How do we know the implementation you suggest is preferable in the general case?

Something that will help answer these questions would be to include a unit test demonstrating the intended behavior and a changelog entry that goes along with this change explaining to the user what to expect differently as a result of this change (you can author it here or include it in CHANGES.rst).

[blueyed]

Thanks for your in depth response!

marker.evaluate should compare markers case-insensitive

Good point. That is not being done in pip's vendorized version (https://github.com/pypa/pip/blob/67f534d3e36bd91e5e8f03615886d200786fb4b4/pip/_vendor/pkg_resources/__init__.py#L2873-L2891), setuptools 19.4.

Have you read pypa/pip#3661 - I could imagine that it would be fixed by updating setuptools / pkg_resources in pip after all.
I've only tried to update the code in that region, but that failed - probably because other parts would have to be updated, too.

[jaraco]

I have read pypa/pip#3661, but it wasn't clear to me from there what the issue was there, and it referenced another issue.

I could imagine that it would be fixed by updating setuptools / pkg_resources in pip after all.

Perhaps, if you believe the current version of setuptools is adequate... but not if you're thinking of diverging the vendored code - that would be undesirable.

[jaraco]

I welcome a follow-up proposal to retain case-insensitivity in the comparison, perhaps by updating the vendored dependency to a new version.

[mindw]

Below is a simple test case which reproduces the issue with hyphens. @blueyed patch fails in this case. This affects the wheel distribution - ed25519ll; extra == "faster-signatures". Perhaps implemeneting extras comparison instead of using normalizaion yield better results?

    def test_marker_evaluation_with_extras_hyphen(self):
        """Extras are also evaluated as markers at resolution time."""
        ad = pkg_resources.Environment([])
        ws = WorkingSet([])
        # Metadata needs to be native strings due to cStringIO behaviour in
        # 2.6, so use str().
        Foo = Distribution.from_filename(
            "/foo_dir/Foo-1.2.dist-info",
            metadata=Metadata(("METADATA", str("Provides-Extra: baz-lightyear\n"
                               "Requires-Dist: quux; extra=='baz-lightyear'")))
        )
        ad.add(Foo)
        assert list(ws.resolve(parse_requirements("Foo"), ad)) == [Foo]
        quux = Distribution.from_filename("/foo_dir/quux-1.0.dist-info")
        ad.add(quux)
        res = list(ws.resolve(parse_requirements("Foo[baz-lightyear]"), ad))
        assert res == [Foo, quux]