Pull Request Check List

Resolves: #issue-number-here

• Added tests for changed code.
• Updated documentation for changed code.

The attempt to groupby during exports can be foiled by cases where a dependency is present with more than one version: eg if we see [foo 1, foo 2, foo 1] then we end up with three groups rather than two.

This improves #5141, in particular the lines for click that we now get for the example given there are

click==7.1.2 ; python_version >= "3.6" and python_full_version < "3.0.0" and python_version < "3.7" or python_version >= "3.6" and python_version < "3.7" and python_full_version >= "3.5.0"
click==8.0.3 ; python_version >= "3.7"

which looks acceptable.

However, in the effort to simplify this to a testcase I ran into at least two bugs that I don't know how to fix, and these are exhibited by the testcase...

Bug 1

Here's what we get for click:

click==7.1.2 ; python_version >= "2.7" and python_full_version < "3.0.0" and python_version < "3.7" or python_full_version >= "3.5.0" and python_version < "3.7" or python_full_version >= "3.6.2" and python_full_version < "4.0.0" or python_version >= "2.7" and python_full_version < "3.0.0" or python_full_version >= "3.5.0"
click==8.0.3 ; python_version >= "3.6"

two of the 7.1.2 possibilities overlap the 8.0.3 marker, which is not OK.

The cause seems to be that when the export code comes to process the "top-level" entries for click-didyoumean and click-plugins it sees that these should be installed for a broad range of python versions, and decides that therefore so should their dependency click.

As best I can see the fundamental problem is that the lock file does not record which packages are actually top-level dependencies, and which are not. I don't see how to fix this without recording the actual top-level requirements as taken from pyproject.toml, which would allow the export code to reach click-didyoumean and click-plugins only in a path where it is taking account of parent markers.

Bug 2

The output markers for click-plugins and click-didyoumean (sort of) are also faulty in the test.

 click-didyoumean==0.0.3 ; python_version < "3.7"
 click-didyoumean==0.3.0 ; python_full_version >= "3.6.2" and python_full_version < "4.0.0"
 click-plugins==1.1.1 ; python_version < "3.7"

• click-plugins should be installed unconditionally
• click-didyoumean 0.0.3 is right, but only accidentally
• click-didyoumean 0.3.0 is wrong, it should be >= "3.7" (per bug 1, it is getting the versions from the "top-level" which is wrong)

The new problem here is that the marker setter at https://github.com/python-poetry/poetry-core/blob/0113a5a172e57424dbf5b2031d8c77097c725b5e/src/poetry/core/packages/dependency.py#L161 does not update the python versions correctly:

• start with a dependency whose marker is "< 3.7"
• update to an Any marker
• that setter does not see python_versions in the Any marker, and therefore does not clear "python < 3.7" from the _python_versions