Re-evaluate use of Poetry in FawltyDeps

[jherland]

The Python community seems to be adopting PEP621 for storing project metadata in pyproject.toml. We have been using Poetry from the start of FawltyDeps, which uses a different Poetry-specific format. Although it's not vitally important for us to follow PEP621, it would be nice to follow best practices in the Python community here.

There is some movement in Poetry to allow support for PEP621 metadata, but we don't know if or when this would be merged:

• Open issue: https://github.com/python-poetry/poetry/issues/3332
• Draft PR: https://github.com/python-poetry/poetry/pull/9135

What are we using Poetry for?

• Creating and entering the development environment for FawltyDeps:
  • From our CONTRIBUTING.md:

    poetry install --with=dev
    poetry shell

  • Our shell.nix perform similar commands to create/enter the Poetry dev shell.
• Setting up CI action environments (most files under .github/workflows/ follow the same pattern):
  • Use Gr1N/setup-poetry@v8
  • Cache the Poetry virtualenv at ~/.cache/pypoetry/virtualenvs, the cache key includes a hash of the poetry.lock file.
  • Use poetry install --no-interaction --sync --with=$DEPENDENCY_GROUPS or similar to setup environment (this is instead of Nox' setup).
  • Use poetry run nox --no-venv --no-install --non-interactive -s ... to run the Nox action.
• Accessing Poetry dependency groups when running Nox sessions:
  • We need to convert the Poetry dependency group that corresponds to a Nox session into a requirements.txt file that can be consumed by Nox's session.install():
  • poetry export --format=requirements.txt --output=.../reqs_from_poetry.txt --with=... or similar
  • The resulting reqs_from_poetry.txt file is passed to Nox' session.install("-r", ...)
• Publishing new releases:
  • Details described in https://github.com/tweag/FawltyDeps/issues/370
  • Store PyPI access token with poetry config pypi-token.pypi <token>
  • Our version number is stored along with the rest of the project metadata in the [tool.poetry] section in pyproject.toml.
  • poetry build and poetry publish accomplishes the actual release.
  • The publish CI action performs similar steps to publish a test release to https://test.pypi.org/project/fawltydeps/

Problems with Poetry?

• Some overlap with Nox?
  • In CI we are setting up the test venv with Poetry instead of Nox. Is this only to avoid duplicate/nested venvs, or is it easier to cache Poetry venvs than Nox venvs?
• Creating/locking the poetry.lock file is sometimes really slow. Have to run with --no-update to prevent unrelated updates to dependencies.
• How to support test scenarios where we want to automatically test FawltyDeps against the oldest/newest set of acceptable dependencies on some platform?
  • Not supported by Poetry AFAICS, but seem to be supported by uv: https://github.com/astral-sh/uv?tab=readme-ov-file#resolution-strategy
• TODO: ISTR that there are more issues that we have with Poetry, but I fail to recall them at this moment.

Alternatives?

• Hatch to build packages and publish to PyPI
  • Part of the PyPA, and thus follows Python packaging standards closely.
  • Hatch also seems to support environment management: https://hatch.pypa.io/latest/environment/
• uv can be used to create/install venvs very quickly.
  • uv only produces resolutions for a single platform and Python version at a time. Unlike Poetry (and PDM), uv does not yet produce a machine-agnostic lockfile. What changes does this mean for our workflow?
• Flit is another PEP621-compatible solution for building and publishing Python packages.

Remaining Questions?

• Does PEP621 support the kind of dependency groups we are currently using in Poetry, and how we are using these from Nox? If not, what is the most appropriate way to express our various Nox session dependencies in PEP621?
• Tangential: What is the optimal cache behavior on GitHub CI?
  • Cache full virtualenvs and re-install as little as possible between runs?
  • Only cache (uninstalled) Python packages and rely on uv to recreate virtualenvs really quickly?
  • Nix??? Currently we're not using Nix in CI at all.
  • Run bechmarks between various cache scenarios.
  • Always try to stay as close to local development as possible. It should be easy to reproduce CI failures locally with Nix and Nox.

[jherland]

An excellent resource for comparing Poetry and similar Python development and packaging-related tools: https://alpopkes.com/posts/python/packaging_tools/#tools-that-do-not-fit-the-categories

(found via the Talk Python to Me podcast: https://www.youtube.com/watch?v=sxN-Iu-okdo)