Should django evaluate typing annotations at all?

[93578237]

I previously opened a bug report #36903 related to deferred annotations in Python 3.14. I believe Django does not need to evaluate annotations using annotationlib.Format.FORWARDREF and could instead use annotationlib.Format.STRING.

[Jacob Walls]

Thanks, but I don't quite follow, doesn't FORWARDREF leave the annotations unevaluated? I think I benched this at the time and found it was faster than STRING. It's also more friendly for any package extending Django's behavior, since it makes fewer assumptions.

[93578237]

from inspect import *; from annotationlib import *; from typing import *

if TYPE_CHECKING: x = int

def foo(a: x) -> None: ...

In [4]: %timeit signature(foo, annotation_format=Format.FORWARDREF)
25 μs ± 445 ns per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

In [5]: %timeit signature(foo, annotation_format=Format.STRING)
12.4 μs ± 495 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

What am I missing?

[Jacob Walls]

Good question. Sorry, I wasn't trying to hide details, I just needed to recall what I benched. I benched the case where the annotations are available (e.g. not guarded under TYPE_CHECKING, which is IMO more likely):

In [13]: class Imported:
    ...:     ...
    ...: 

In [14]: def foo(a: Imported, b: Imported, c: Imported): ...

In [15]: %timeit signature(foo, annotation_format=Format.FORWARDREF)
4.52 μs ± 177 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

In [16]: %timeit signature(foo, annotation_format=Format.STRING)
7.59 μs ± 37.3 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)