Support the free-threaded build of CPython 3.13

[lysnikolaou]

Is your feature request related to a problem?

No response

Describe the solution you'd like

Hey all! 👋

I took the liberty of looking into adding support for the 3.13 free-threaded build of CPython after multidict (I also plan to look at other libraries under the aio-libs organization). This aims to be a tracking issue for all work necessary to support it, for which I'm going to open a PR shortly.

• Set up CI for free-threading
• Audit the Cython extension module for thread-safety issues
  • The only issues seems to be the usage of a statically allocated buffer for the Writer struct. That's fairly straightforward to solve.
• Audit the Python code for thread-safety issues
  • One fairly complicated thing to figure out was that there's some thread safety issues in the Python code as well. This stems from havily using non-thread-local caches when creating URL instances. Instantiating a URL with a string url for example calls encode_url which uses an lru_cache. This means that multiple threads might end up sharing the same object, even if the URL object appears to have been created locally. There's a couple possible solution to this:
    • Disable lru_cache under the free-threaded build so that object sharing between threads is more straightfoward.
    • Instead of using lru_cache to cache the URL object itself, it could instead cache the immutable parts that make up a URL.
• Mark the Cython extension module as thread-safe with setting freethreading_compatible in the cython directives (can be easily done through pyproject.toml).
• Build wheels for the free-threaded build.

Hope this all sounds good! Let me know in case there's something that doesn't seem right.

Describe alternatives you've considered

• We could say that people testing on the free-threaded build should only install the pure-Python wheels, but that doesn't seem like the best say forward, since the required changes are not that intrusive.

Additional context

No response

Code of Conduct

• I agree to follow the aio-libs Code of Conduct