Skip to content

ci: migrate 13 workflows to needs.setup.outputs.matrix (batch 8/8) - #4917

Open
luhenry wants to merge 13 commits into
mainfrom
fix/matrix-batch-8
Open

luhenry wants to merge 13 commits into
mainfrom
fix/matrix-batch-8

Conversation

@luhenry

@luhenry luhenry commented Nov 18, 2026 •

Copy link
Copy Markdown
Member

Part of the repo-wide migration off hardcoded strategy.matrix python/libc arrays
onto the setup job's generated outputs.matrix (see CLAUDE.md / PRs #4914, #4915).

Packages in this batch, each as its own commit:

  • thinc
  • thriftpy2
  • traits
  • turbohtml
  • type-enforced (type_enforced)
  • ulid-transform
  • unicodedata2
  • vhacdx
  • vtk
  • webrtcvad-wheels
  • yara-python
  • zlib-ng
  • zttp

For each file: the build_wheels job's strategy.matrix now reads
include: ${{ fromJSON(needs.setup.outputs.matrix) }} instead of its own
hardcoded version/python(/libc) arrays, and every remaining
${{ matrix.libc }} reference (job name:, CIBW_BUILD/only:,
CIBW_BEFORE_ALL_LINUX, artifact name:) was renamed to
${{ matrix.platform }}. matrix.python and matrix.version are unchanged.
None of the publish jobs referenced matrix.libc.

Pre-existing stale/incomplete matrix issues found (fixed as a side effect of
this structural change):

  • Every file in this batch except turbohtml and type-enforced had a
    hardcoded python: list that starts at cp312, while
    docs/packages/<pkg>.yaml's files: entries for the currently pending
    version(s) also call for cp310/cp311 (and, where applicable,
    cp315/cp315t). Those older/newer interpreters were never being built.
    (turbohtml and type-enforced currently have no pending version, but
    their hardcoded lists exactly match the latest released version's
    files:, so there's nothing to fix there.)
  • thinc: the old job hardcoded python: ["cp312","cp313","cp314"] with a
    manual exclude: dropping cp313/cp314 for 9.1.1. 9.1.1's own
    files: list in docs/packages/thinc.yaml never declared cp313/cp314
    in the first place, so the generated matrix naturally omits them - the
    manual exclude was redundant and is dropped. Separately,
    docs/packages/thinc.yaml declares musllinux files for both pending
    versions (9.1.1 and 10.0.0) that the old manylinux-only job never built;
    this PR starts building those for the first time (added
    MUSLLINUX_RISCV64_IMAGE/CIBW_MUSLLINUX_RISCV64_IMAGE, parameterized the
    job name/only:/artifact name on ${{ matrix.platform }}, and widened the
    publish artifact-pattern from *-manylinux_riscv64 to *riscv64).
  • webrtcvad-wheels and yara-python: the old hardcoded lists included
    cp314/cp314t, but neither package's currently pending version actually
    declares cp314/cp314t files in its docs/packages/*.yaml (upstream
    doesn't publish those wheels for these older releases) - those legs were
    being built unnecessarily/incorrectly. The generated matrix now builds
    exactly what's declared.
  • vtk: docs/packages/vtk.yaml has no musllinux entries for any pending
    version, so this one stays manylinux-only; only the strategy.matrix
    block itself changes. Flagging per the task note that VTK's build is
    unusually large/complex (48h timeout, custom CMake/podman build, a
    same-leg SDK-packaging step gated on matrix.python == 'cp312') - I
    verified the matrix shape is otherwise standard (just version +
    python, no libc, same as thinc) and found nothing non-standard
    beyond that, but called it out explicitly rather than guessing given its
    size.

No changes to docs/packages/*.yaml, .queue.yml, or skills/ in this branch.
Not waiting on this PR's CI to go green before reporting back.

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