aio-libs / multidict / 36554860349

87%

push

github

Build every tested dist in one stage, test the sdist (#1638)

<!-- Thank you for your contribution! -->

## What do these changes do?

The `test` matrix needed the whole `build-wheels-for-tested-arches`
matrix, so on the last master run its legs started at +8.4 min, held up
by the macOS wheel job, although the only wheels they install were ready
at +3.8. Those Linux wheels now come from `build-linux-wheels`, a
stage-1 job built from the sdist that `cibuildwheel` does not test
(`CIBW_TEST_SKIP=*-manylinux_x86_64`), because the test matrix runs the
same suite on them with coverage. The i686 wheels, built at release time
only, keep their `cibuildwheel` test run since nothing else covers them.

What is left goes to `build-wheels-for-other-platforms` (macOS, Windows
and, on tags, musllinux), where `cibuildwheel` tests each wheel itself.
It no longer gates the tests; it gates `test-summary`, so a broken wheel
still fails the run.

`test-cython-capi`, `inlining`, `hypothesis-gil`, `hypothesis-ft`,
`asan` and `tsan` now take their source from the sdist, the way `test`
already did, so a file missing from the release tarball fails CI rather
than a user's install. `MANIFEST.in` grafts `tools/` for the inlining
check. `benchmark` keeps its checkout: the CodSpeed runner opens the
workspace as a git repository and fails without one. Apart from it,
`lint` and `codeql`, `build-pure-python-dists` is the only job that
checks out the repository.

Two conditionals changed. The `windows-11-arm` exclude never fired:
`cond && '' || 'none'` always yields `'none'`, because `''` is falsy, so
that wheel was built on every run instead of on release tags. And x86_64
macOS wheels are tested under Rosetta at about twice the arm64 cost, so
pull requests now build them without testing them, while master, the
merge queue and tags test them as before.

The `test` matrix also drops its single-valued `os` dimension.

## Are there changes in behavior for the user?

No, this is a CI-only c... (continued)

788 of 1576 branches covered (50.0%)

Branch coverage included in aggregate %.

8961 of 9638 relevant lines covered (92.98%)

1.86 hits per line