gh-138764: annotationlib - Check if the VALUE_WITH_FAKE_GLOBALS format is implemented before calling an __annotate__ function with empty fake globals

[DavidCEllis]

This fixes an issue where user defined __annotate__ functions that do not support a specific format would be called with fake globals in a way that they couldn't correctly raise NotImplementedError.

One downside is that some annotate functions that may have appeared to 'work' before now correctly fail. It's possible we may want to add some kind of fallback to VALUE annotations in these cases, but calling with fake globals as it does now would only accidentally succeed depending on exactly how the function performs the comparisons.

from annotationlib import Format, call_annotate_function
def annotate(format, /):
    if format == Format.VALUE:
        return {'x': str}
    else:
        raise NotImplementedError(format)

print(call_annotate_function(annotate, Format.STRING))

Current:

{'x': 'str'}

With patch:

NotImplementedError: 2

But change the comparison to check if the format is in a set:

from annotationlib import Format, call_annotate_function

def annotate(format, /):
    if format in {Format.VALUE}:
        return {'x': str}
    else:
        raise NotImplementedError(format)

print(call_annotate_function(annotate, Format.STRING))

Current:

TypeError: exceptions must derive from BaseException

With patch:

NotImplementedError: 2

• Issue: annotationlib - call_annotate_function should check VALUE_WITH_FAKE_GLOBALS is implemented before calling in a namespace with empty fake globals #138764

[JelleZijlstra]

Not a fan of these tests as they're asserting details of the implementation, not the observable behavior. Is there a way to write a test that's closer to your use case?

[DavidCEllis]

I suppose this could instead test multiple annotate functions that provide only those formats?

Let me know if you prefer this or #138803 and I'll modify the tests for the relevant PR.

[DavidCEllis]

Closing in favour of the approach in #138803