🐛 Use inspect to determine whether the AnyIO version supports abandon_on_cancel

[Stealthii]

#202 introduces an issue on py38 and py39 with an earlier version of importlib.metadata not detecting versions in certain virtual environments. It also uses string comparisons:

Traceback (most recent call last):
  File "example/lib64/python3.8/site-packages/asyncer/__init__.py", line 3, in <module>
    from ._main import PendingValueException as PendingValueException
  File "example/lib64/python3.8/site-packages/asyncer/_main.py", line 17, in <module>
    from asyncer._compat import run_sync
  File "example/lib64/python3.8/site-packages/asyncer/_compat.py", line 16, in <module>
    if ANYIO_VERSION >= "4.1.0":
TypeError: '>=' not supported between instances of 'NoneType' and 'str'

This PR addresses the issue by using inspect to determine anyio's supported kwargs for the run_sync method.

[Stealthii]

This bug affects v0.0.8 (the latest version) only.

The PR also contains as prior solution to correctly detect AnyIO package version using importlib-metadata and numeric comparison for Python < 3.10, however this was ultimately replaced with the kwargs check.

We should avoid relying on importlib.metadata outside of build environments as it relies heavily on discoverable dist-info or egg-info metadata, which may not be available in non-virtualenv setups (such as system-level packages).

[Stealthii]

@tiangolo I would recommend merging and releasing this in the next version before dropping Python 3.8 support.

[Stealthii]

This still affects non-venv environments. This should be merged and released, as any packages using this that are installed as part of a system install (or flat package) have to explicitly block v0.0.8 from dependency mapping.

[auvipy]

can you check why the check level is failing?

[svlandeg]

Hi @Stealthii, thanks for the contribution, and apologies for the delay in reviewing this!

Comparing the ANYIO_VERSION with a string comparison is definitely a bug, and something we'll make sure to avoid in the future.

I think it makes sense to use a direct inspect call to check the availability of the correct argument. All the changes in this PR look good to me, and it even removes some duplicate code which is great 💯

[tiangolo]

Thank you! 🚀

There's one small issue with the current implementation, the check for abandon_on_cancel will be executed every time run_sync is called. With something closer to the previous implementation (with some duplicated code) the check is run only once when declaring the function, but after that, one of the two versions of the function will be the only one declared and executed.

Maybe we could move the if statement to the position where the previous (and problematic) check was.

[Stealthii]

Agreed @tiangolo - oversight on my part when throwing that into the ring when rebasing recently!

I can revert 78626db, or perhaps move the kwarg check to a constant on import? (done in 0227077). Don't mind either way, just conscious I might not be matching codestyle on this library.

[tiangolo]

Thanks @Stealthii for the work (and the patience).

I can revert 78626db, or perhaps move the kwarg check to a constant on import? (done in 0227077).

I'm not sure I fully got this part 😅

But to simplify, I would leave the code almost as it currently is before this PR, trading:

if ANYIO_VERSION >= "4.1.0":

for:

if RUN_SYNC_AOC:

Don't worry too much about the commit history, in the end I would squash and merge.

[svlandeg]

@tiangolo: You're right - the code duplication is useful in this case because it prevents the check being run every time the function is executed. I've restored the code to have the same structure as current main, but now using the new RUN_SYNC_AOC bool.

[Stealthii]

@tiangolo: You're right - the code duplication is useful in this case because it prevents the check being run every time the function is executed. I've restored the code to have the same structure as current main, but now using the new RUN_SYNC_AOC bool.

The original reason I moved the check into a constant (once when importing the module), and had if RUN_SYNC_AOC inside the function was type checkers struggling with run_sync being defined on a conditional, and not inspecting things correctly. It's been over a year since I looked at this issue, so I'm unsure if this is still an issue.

because it prevents the check being run every time the function is executed

Checking a boolean isn't that expensive honestly.

[Stealthii]

This is still an issue, and with the recently released 0.0.9 having the bug too, it's causing the need for build systems to blocklist both 0.0.8 and 0.0.9 now from dependency constraints.

Traceback (most recent call last):
  File "example/lib/python3.11/site-packages/asyncer/__init__.py", line 3, in <module>
    from ._main import PendingValueException as PendingValueException
  File "example/lib/python3.11/site-packages/asyncer/_main.py", line 17, in <module>
    from asyncer._compat import run_sync
  File "example/lib/python3.11/site-packages/asyncer/_compat.py", line 16, in <module>
    if ANYIO_VERSION >= "4.1.0":
       ^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: '>=' not supported between instances of 'NoneType' and 'str'