Build in separate jobs. This makes it clearer if e.g. a single Python
version is failing. It also potentially gets you more parallelism.
Build everything on push to master.
Only build Linux 3.8 and 3.11 wheels on PRs.
This implements PEP 621, obviating the need for `setup.py`, `setup.cfg`,
and `MANIFEST.in`. The build backend Hatchling (of which I am a
maintainer in the PyPA) is now used as that is the default in the
official Python packaging tutorial. Hatchling is available on all the
major distribution channels such as Debian, Fedora, and many more.
## Python support
The earliest supported Python 3 version of Hatchling is 3.7, therefore
I've also set that as the minimum here. Python 3.6 is EOL and other
build backends like flit-core and setuptools also dropped support.
Python 3.6 accounted for 3-4% of downloads in the last month.
## Plugins
Configuration is now completely static with the help of 3 plugins:
### Readme
hynek's hatch-fancy-pypi-readme allows for the dynamic construction of
the readme which was previously coded up in `setup.py`. Now it's simply:
```toml
[tool.hatch.metadata.hooks.fancy-pypi-readme]
content-type = "text/markdown"
fragments = [
{ path = "README.md" },
{ path = "CHANGES.md" },
]
```
### Versioning
hatch-vcs is currently just a wrapper around setuptools-scm (which
despite the legacy naming is actually now decoupled from setuptools):
```toml
[tool.hatch.version]
source = "vcs"
[tool.hatch.build.hooks.vcs]
version-file = "src/_black_version.py"
template = '''
version = "{version}"
'''
```
### mypyc
hatch-mypyc offers many benefits over the existing approach:
- No need to manually select files for inclusion
- Avoids the need for the current CI workaround for https://github.com/mypyc/mypyc/issues/946
- Intermediate artifacts (like `build/`) from setuptools and mypyc
itself no longer clutter the project directory
- Runtime dependencies required at build time no longer need to be
manually redeclared as this is a built-in option of Hatchling
Co-authored-by: Richard Si <63936253+ichard26@users.noreply.github.com>
Co-authored-by: Jelle Zijlstra <jelle.zijlstra@gmail.com>
Bumps cibuildwheel from 2.8.1 to 2.10.0 which has 3.11 building enabled
by default. Unfortunately mypyc errors out on 3.11:
src/black/files.py:29:9: error: Name "tomllib" already defined (by an import) [no-redef]
... so we have to also hide the fallback import of tomli on older 3.11
alphas from mypy[c].
We've decided to a) convert stable back into a branch and b) to update
it immediately as part of the release process. We may as well automate
it. And about going back to a branch ...
Git tags are not the right tool, at all[^1]. They come with the
expectation that they will never change. Things will not work as
expected if they do change, doubly so if they change regularly. Once
you pull stable from the remote and it's copied in your local
repository, no matter how many times you run git pull you'll never see
it get updated automatically. Your only recourse is to delete the tag
via `git tag -d stable` before pulling.
This gets annoying really quickly since stable is supposed to be the
solution for folks "who want to move along as Black developers deem
the newest version reliable."[^2] See this comment for how this impacts
users using our Vim plugin[^3]. It also affects us developers[^4]. If
you have stable locally, once we cut a new release and update the stable
tag, a simple `git pull` / `git fetch` will not pull down the updated
stable tag. Unless you remember to delete stable before pulling, stable
will become stale and useless.
You can argue this is a good thing ("people should explicitly opt into
updating stable"), but IMO it does not match user expectations nor
developer expectations[^5]. Especially since not all our integrations
that use stable are bound by this security measure, for example our
GitHub Action (since it does a clean fetch of the repository every time
it's used). I believe consistency would be good here.
Finally, ever since we switched to a tag, we've been facing issues with
ReadTheDocs not picking up updates to stable unless we force a rebuild.
The initial rebuild on the stable update just pulls the commit the tag
previously pointed to. I'm not sure if switching back to a branch will
fix this, but I'd wager it will.
[^1]: https://git-scm.com/docs/git-tag#_on_re_tagging
[^2]: https://black.readthedocs.io/en/stable/contributing/release_process.html#moving-the-stable-tag
[^3]: https://github.com/psf/black/issues/2503#issuecomment-1196357379
[^4]: In fairness, most folks working on Black probably don't use the
`stable` ref anyway, especially us maintainers who'd know what is
the latest version by heart, but it'd still be nice to make it
usable for local dev though.
[^5]: Also what benefit does a `stable` ref have over explicit version
tags like `22.6.0`? If you're going to opt into some odd pin
mechanism, might as well use explicit version tags for clarity
and consistency.
Draft releases don't trigger the workflows (that's good!) but since they only
Commit history before merge:
* fix: run pypi upload from published draft releases
* Fix broken task list markup in PR template
* change docker workflow to build on release publish
Co-authored-by: Richard Si <63936253+ichard26@users.noreply.github.com>
