Shared checks.yml in this repo with a derived Python matrix #95

Activity

1. lesnik512

   added

   enhancementNew feature or requestNew feature or request

   ready-for-agentFully specified, ready for an AFK agentFully specified, ready for an AFK agent

   on Sep 12, 2026
2. lesnik512 commented on Sep 12, 2026

   

   lesnik512

   on Sep 12, 2026

   MemberAuthor

   More actions

   Input from #91 (adding 3.14t to the per-repo _checks.yml): the derived matrix needs two per-entry knobs that the shared workflow has to expose, otherwise two repos go red the day it switches on.

   • Job-level env that applies only to the free-threaded entry. db-retry sets DISABLE_SQLALCHEMY_CEXT_RUNTIME=1 on 3.14t (SQLAlchemy 2.0.x cyext re-enables the GIL; fixed upstream for 2.1), faststream-concurrent-aiokafka sets AIOKAFKA_NO_EXTENSIONS=1 (no cp314t wheel; the C codecs re-enable the GIL). Both use ${{ matrix.python-version == '3.14t' && '1' || '' }} today, so an env input (or a free-threaded-env map) that the workflow spreads onto the pytest job covers it.
   • A way to drop the free-threaded entry per repo. modern-di-arq (hiredis, free-threading support redis/hiredis-py#229) and modern-di-grpc (grpcio, Support no-GIL (free-threaded) Python grpc/grpc#38762) cannot keep the GIL disabled because of a dependency and carry exemption rows on the standard. A boolean input like free-threaded: false is the minimum; the exemption row is the record that justifies setting it.

   Also worth carrying over: the 3.14t entry in every migrated repo runs python -c "import sys, <package>; assert not sys._is_gil_enabled()" before the suite. Pytest only emits a RuntimeWarning when an import re-enables the GIL, so without that step a GIL-enabled run still goes green.
3. lesnik512 commented on Sep 15, 2026

   

   lesnik512

   on Sep 15, 2026

   MemberAuthor

   More actions

   Closing without migrating. #101 built the workflow as designed and db-retry#46 proved it end to end from a branch ref, including the service-container path. The decision is against it, on the shape of the interface rather than on whether it works.

   Why not

   workflow_call inputs are scalars. Anything structured a repo varies, a service container or an env map, has to be flattened into a string: first a health command plus a baked-in POSTGRES_PASSWORD, then raw docker create options. Both are image knowledge leaking into the org repo, and neither reads well. That is a GitHub limit, not a design slip, so it does not go away with better input names.

   What sharing buys is real but bounded: the run-time matrix (the only part section 6 needs) and one-PR policy changes instead of a per-repo sweep. Sweeps are how the org already works, so that saving does not pay for the interface.

   The idea, if this is ever reopened

   A reusable workflow with zero inputs, everything read from the repo:

   • services: the pytest job runs just services-up when just --summary lists it; the recipe is docker compose up -d --wait <service>, so image, env and healthcheck stay in the compose file the repo already keeps for local dev (modern-di-arq has this today as redis-up)
   • test env: tests default to localhost, as three of the five service repos already do
   • free-threaded exemption: derived from the Free Threading classifier section 11 requires; today exactly the two exempt repos lack it
   • docs job: when the justfile has docs-build
   • the t-build switch two repos need (DISABLE_SQLALCHEMY_CEXT_RUNTIME, AIOKAFKA_NO_EXTENSIONS): exported by test-ci when the interpreter reports the GIL disabled

   Every caller is then the same two lines. The matrix script and its tests in #101 are the reusable part of that work.

   Section 7 of the standard still says the shared checks.yml "will live" here; that sentence needs its own change now.
4. lesnik512

   closed this as not plannedon Sep 15, 2026