I've been experimenting with using pip-compile for Jinja, and have run into an interesting circular dependency between Jinja and Sphinx. Jinja uses an editable install (-e .) for development. Jinja lists Sphinx in requirements.in for building docs. But Sphinx depends on Jinja, so jinja2==2.11.1 gets pinned in requirements.txt. But this pin overwrites the locally installed version of Jinja, which is what should be documented and tested.

I wanted to provide contributors with a command to set up a dev env. Using the current resolver, pip install -e . -r requirements.txt ends up installing jinja2==2.11.1 from requirements.txt instead of -e . The order of -e and -r in the command doesn't matter.

Using the alpha resolver, pip install --unstable-feature=resolver -e . -r requirements.txt (or -r -e) produces an error.

$ pip install --unstable-feature=resolver -r requirements/dev.txt -e .
Obtaining file:///home/david/Projects/jinja
ERROR: Could not find a version that satisfies the requirement ExplicitRequirement(EditableCandidate('file:///home/david/Projects/jinja'))
ERROR: Could not find a version that satisfies the requirement jinja2==2.11.2
ERROR: No matching distribution found for jinja2, jinja2

I would expect a local install to always take precedence over pinned requirements. Otherwise, projects that depend on each other during development will have a more difficult time pinning requirements.

As a workaround, the editable install can be done as a separate command. Tox, Nox, and ReadTheDocs can install the local project after the dependencies, but adding a second command introduces a hard to detect way for new contributors to misconfigure their local environment.