TypeError: TypedDict does not support instance and class checks

[KyleKing]

I saw in #7 that TypedDict should be supported from 0.7.0 and partial support has been in beartype since 0.1.0, but I'm running into errors with Python 3.7.12 and 3.9.7, so I'm not sure if this snippet should be working:

# beartype==0.8.1

from beartype import beartype

try:
    from typing import TypedDict
except ImportError:  # For Python < 3.8
    from typing_extensions import TypedDict  # typing-extensions = "^3.10.0"

class Point2D(TypedDict):
    x: int
    y: int
    label: str

@beartype
def convert(point: Point2D) -> Point2D:
    return point

convert({'x': 1, 'y': 2, 'label': 'good'})

> python tmp.py                                                                                                                    ─╯
Traceback (most recent call last):
  File "/Users/kyleking/.pyenv/versions/3.9.7/lib/python3.9/site-packages/beartype/_util/cls/utilclstest.py", line 159, in die_unless_type_isinstanceable
    isinstance(None, cls)  # type: ignore[arg-type]
  File "/Users/kyleking/.pyenv/versions/3.9.7/lib/python3.9/typing.py", line 1956, in __subclasscheck__
    raise TypeError('TypedDict does not support instance and class checks')
TypeError: TypedDict does not support instance and class checks

Should beartype properly check TypedDicts? There were similar errors found in Pydantic and some notes here. TypedDict can't be used with isinstance by design. There is a reference to tests for TypedDict, but they don't appear to be present

And this isn't promising 😅

beartype/beartype/_data/hint/pep/sign/datapepsigns.py

Line 131 in b7ab8b0

# is_typeddict <-- unusable as a type hint

[leycec]

...ohnoes. You caught us red-handed – or honey-pawed, as it were. ^{badum ching}

A long time ago in a codebase far, far away, @beartype once shallowly type-checked typing.TypedDict... I think, anyway. Tragically, our unavoidable laziness coupled with aggressive release schedules means we kinda failed to properly test that.

Welp, this is awkward. Some unspecified recent release of beartype (I am glaring at you, beartype 0.8.0) silently regressed that and broke both typing.TypedDict and typing.NamedTuple. I am now shaking my emaciated fists at my younger self, on whose behalf I now extend my most grievous apologies.

Thanks so much for the compassionate assistance, Master Kyle. Let's get this patched up for everyone and pretend this never happened! The sweat pouring off my Klingon-like forehead is real. 😓 💦

[leycec]

Welp, this is even more embarrassing...

Turns out we actually never supported typing.TypedDict at all! Our documentation claiming we supported typing.TypedDict shallow type-checking from beartype 0.1.0 onward was, in fact, complete balls.

On the bright side, at least we now know why a non-existent feature was never tested.

This is a pretty critical omission, as typing.TypedDict is a core type. So, I've slated this for inclusion in the upcoming beartype 0.9.0 release. Let's optimistically pretend that I'll get everything pushed out this Friday. 🤕

[KyleKing]

Thanks! I'm always impressed by your liberal use of puns and references

I would be happy to test and give feedback before you release, if that would help. And no rush!

[leycec]

Oh, thank you so much. You are too kind and complimentary. This elicits feelings of warmth and generosity, which means I now want to get this done even more! Nice-guy Worf approves.

Please do liberally test this after beartype 0.9.0 drops and complain with clenched fists here if we lied yet again about this working.

...if we type-check it, they will come

In theory, it should be trivial for beartype to at least shallowly type-check typing.TypedDict subclasses by just internally reducing the typing.TypedDict superclass on subclass types to the dict superclass at decoration time: e.g.,

# Let's just pretend this...
from typing import TypedDict
class Point2D(TypedDict):
    x: int
    y: int
    label: str

# ...was coded like this, instead!
class Point2D(dict):
    x: int
    y: int
    label: str

wut you gonna do

TypeDict Rant, Let's Go!

Lastly, I kinda agree with mpkocher's well-written typing.TypedDict tl;dr:

I believe TypedDict can be a value tool to help improve clarity of interfaces, specifically in legacy code that is a bit dictionary-mania heavy. However, for new code, I would suggest avoid using TypedDict in favor of the thin data models, such as pydantic and attrs [and beartype dataclasses – some day over the GitHub rainbow when I gather the internal strength to implement them].

typing.TypedDict occupies this lonely weird liminal space in Python's type-hinting landscape, where:

• It doesn't really behave like anything else that's been standardized (even typing.NamedTuple, which behaves much more sensibly).
• Breaks usability by preventing user-defined subclasses and instances from being compared with other arbitrary classes (e.g., by being passed to the isinstance() or issubclass() builtins).
• Fails to interoperate cleanly at runtime with the rest of the Python ecosystem for aforementioned reasons.

It's a crying shame. Guido & Co. were clearly trying to extend the idea of fixed-length tuple type hints (e.g., typing.Tuple[int, str, bool]) to dictionaries. Sadly, they fumbled the ball in their own end zone after tripping over the marching band furry. Now, everyone just uses pydantic instead.

I am sighing.

[leycec]

Brutally resolved! In commit 342e5fc (and the upcoming beartype 0.9.0 – it's any day now, we swear it on Bearzabubba), @beartype now shallowly type-checks both typing.TypedDict and typing_extensions.TypedDict type hints as semantically equivalent to collections.abc.Mapping[str, object].

Clearly, this is... suboptimal. But we can agree that something beats nothing. Well, conditionally depending on the exact nature of that something. So sometimes. And this is one of those times.

Thanks again for the thoughtful heads-up and bearing ^{hyuk hyuk} with us here, Kyle the undisputed King of Parexel AI Labs. You deserve all the academic citations that the @beartype community has to give. Reference Kyle, everyone! He rocks that AI. 📝