Hello,
Currently, for macOS we're publishing both arm64, x86_64 *and*
universal2 binary wheels (the latter contain both arm64 and x86_64 code
in a single binary).
Here are some observations from me:
* Producing universal2 wheels is more complex than producing
single-architecture wheels (we actually have to build for both
architectures separately, then merge the results); it's also one more
CI/packaging configuration to look after
* x86-64 Macs are legacy and are gradually disappearing (especially for
high-performance applications where ARM Macs are massively faster)
* Numpy publishes only architecture-specific wheels, while Pandas
publishes both architecture-specific and universal wheels
* Size-wise, a universal wheel is not much smaller than the sum of
architecture-specific wheels (for example, 43.7 MB for
pyarrow-9.0.0-cp310-cp310, vs. 24.0 MB + 21.6 MB)
Is there any reason why we should continue publishing universal wheels
for macOS?
Regards
Antoine.
