Skip to content

ci: migrate 14 workflows to needs.setup.outputs.matrix (batch 1/8) - #4920

Open
luhenry wants to merge 13 commits into
mainfrom
fix/matrix-batch-1
Open

luhenry wants to merge 13 commits into
mainfrom
fix/matrix-batch-1

Conversation

@luhenry

@luhenry luhenry commented Nov 18, 2026 •

Copy link
Copy Markdown
Member

Replaces the hardcoded python:/libc: matrix arrays in build_wheels (or
the equivalent build job) with include: ${{ fromJSON(needs.setup.outputs.matrix) }},
and renames the remaining matrix.libc references to matrix.platform to
match the generated matrix's field names. matrix.python and matrix.version
are untouched.

Packages touched (14):

  • aiocsv
  • aiofastnet
  • annotatedyaml
  • apache-beam
  • atomic-dict
  • bitstruct
  • can-ada
  • ckzg
  • clickhouse-cityhash
  • cmarkgfm
  • coincurve
  • componentize-py
  • convertertools

Skipped

  • audiostretchy — not converted. Its build_wheels job builds a single
    universal py3-none wheel (hardcoded only: cp312-manylinux_riscv64, no
    matrix.python/matrix.libc at all) and its separate test_wheel job has
    its own hand-picked, build-unrelated test matrix (python: ["cp313", "cp314", "cp314t"], deliberately excluding cp312 since that interpreter
    is already exercised by the build). Neither shape matches what
    needs.setup.outputs.matrix represents (one entry per build
    python/platform combo), and test_wheel's list already includes cp314t,
    which isn't even a declared target in docs/packages/audiostretchy.yaml.
    Converting it would change test coverage in ways outside this mechanical
    refactor's scope, so I left it as-is.

Pre-existing stale-matrix mismatches found (step 4)

Verified each file's old hardcoded python/libc list against uv run --quiet ci_scripts/pending_versions.py --python-matrix --python-exclude cp39 <pkg> for the currently pending version(s). Several files had drifted out of
sync with docs/packages/<pkg>.yaml, same class of bug as #4914/#4915. The
structural fix in this PR corrects these as a side effect:

  • apache-beam: old hardcoded python: [cp312, cp313, cp314] (with a
    per-version exclude for 2.61.0 dropping cp313/cp314) never built cp310 or
    cp311 at all. docs/packages/apache-beam.yaml declares cp310/cp311 targets
    for every pending version (2.61.0, 2.75.0, 2.76.0, 2.77.0) — this PR starts
    building them.
  • atomic-dict: old hardcoded python: [cp312, cp313, cp314, cp314t] x libc: [manylinux, musllinux] both built orphan combos nothing in
    docs/packages/atomic-dict.yaml declares (cp312 and cp314t, for either
    platform) and never built the cp310-manylinux wheel 0.5.0 actually needs.
  • bitstruct: old hardcoded python: [cp312, cp313, cp314, cp314t] never
    built cp310/cp311, which both pending versions (8.23.0, 8.24.0) declare, nor
    cp315/cp315t, which both versions now also declare.
  • ckzg: old hardcoded python: [cp312, cp313, cp314, cp314t] never built
    cp310/cp311, which the pending version (2.1.8) declares.
  • clickhouse-cityhash: old hardcoded python: [cp312, cp313, cp314]
    never built cp310/cp311, which the pending version (1.0.2.6) declares.
  • cmarkgfm: old hardcoded python: [cp312, cp313, cp314, cp314t] never
    built cp310/cp311, which the pending version (2025.10.22) declares.
  • coincurve: old hardcoded python: [cp312, cp313, cp314] never built
    cp310/cp311, which the pending version (21.0.0) declares; it also built an
    orphan cp314 combo that docs/packages/coincurve.yaml never declares for
    21.0.0 (no wheel for it exists to publish), which this PR stops building.
  • convertertools: old hardcoded python: [cp312, cp313, cp314, cp314t]
    never built cp311, which the pending version (1.1.0) declares.

No mismatch was detectable for aiocsv or annotatedyaml — both
currently have no pending version (pending_versions.py returns []), so
there's nothing to compare against right now; the conversion is still
structurally correct for whenever a new version lands.

Exclusion preserved (step 5)

  • can-ada: the file's own comment ("Upstream builds cp39-cp314, no
    free-threaded wheel; floor is our own cp312") documents a deliberate choice
    to not build cp310/cp311, beyond the standard python-exclude: 'cp39'
    default. docs/packages/can-ada.yaml does declare cp310/cp311 targets for
    the pending version (3.0.0), so without an explicit exclude this PR's
    conversion would silently start building interpreters the port's own
    comment says it doesn't want. Added python-exclude: 'cp39 cp310 cp311' to
    the setup: job's with: block to preserve that decision.

No other file in this batch had a comment indicating a deliberate
cp310/cp311 (or similar) exclusion — grepped all 14 files for floor,
cp310, cp311, exclude to confirm.

Old hardcoded python: [cp312, cp313, cp314, cp314t] x libc: [manylinux,
musllinux] never matched docs/packages/atomic-dict.yaml: it built orphan
cp312/cp314t combos nothing publishes, while never building the cp310
manylinux wheel 0.5.0 actually needs. The matrix conversion fixes this as a
side effect.
Old hardcoded python: [cp312, cp313, cp314, cp314t] never built cp310/cp311,
which docs/packages/bitstruct.yaml declares for both pending versions, nor
cp315/cp315t, which it also now declares.
Also starts building cp310/cp311 wheels for every pending version, which the
old hardcoded python: [cp312, cp313, cp314] list never attempted despite
docs/packages/apache-beam.yaml declaring them (same class of stale-matrix bug
as #4914/#4915).
Adds python-exclude: 'cp39 cp310 cp311' to setup to preserve the file's own
deliberate cp312 floor (docs/packages/can-ada.yaml declares cp310/cp311
targets that the generated matrix would otherwise start building).

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant