When Cython 3.1 is out, we can try using the limited API, which would let us publish many fewer wheels (108 for pyzmq 26.1).
My thought is to perhaps go with some version of SPEC0 for 'full' support:
only build one limited API wheel for lowest supported Python with lowest supported glibc (currently 3.7, manylinux 2010, I think)
don't build any wheels for other Pythons not supported by SPEC0 - this would mean no more wheels needed for CPython 3.8-3.10
possibly build a 3.12 limited wheel for performance reasons (ref)
This would reduce the CPython builds from 8 supported CPythons times ~12 targets to just 5 (2x 3.13).
If there are no downsides to limited wheels after 3.12, we could stop building wheels for each minor version altogether (hopefully free-threading will have a limited API again in 3.14).
If we went with just limited API wheels, we'd only have to build for ~3 CPythons (3.7, 3.12, 3.13t).
When Cython 3.1 is out, we can try using the limited API, which would let us publish many fewer wheels (108 for pyzmq 26.1).
My thought is to perhaps go with some version of SPEC0 for 'full' support:
This would reduce the CPython builds from 8 supported CPythons times ~12 targets to just 5 (2x 3.13).
If there are no downsides to limited wheels after 3.12, we could stop building wheels for each minor version altogether (hopefully free-threading will have a limited API again in 3.14).
If we went with just limited API wheels, we'd only have to build for ~3 CPythons (3.7, 3.12, 3.13t).