|
| 1 | +# Callables as descriptors? |
| 2 | + |
| 3 | +<!-- blacken-docs:off --> |
| 4 | + |
| 5 | +```toml |
| 6 | +[environment] |
| 7 | +python-version = "3.14" |
| 8 | +``` |
| 9 | + |
| 10 | +## Introduction |
| 11 | + |
| 12 | +Some common callable objects (functions, lambdas) are also bound-method descriptors. That is, they |
| 13 | +have a `__get__` method which returns a bound-method object that binds the receiver instance to the |
| 14 | +first argument (and thus the bound-method object has a different signature, lacking the first |
| 15 | +argument): |
| 16 | + |
| 17 | +```py |
| 18 | +from ty_extensions import CallableTypeOf |
| 19 | +from typing import Callable |
| 20 | + |
| 21 | +class C1: |
| 22 | + def method(self: C1, x: int) -> str: |
| 23 | + return str(x) |
| 24 | + |
| 25 | +def _( |
| 26 | + accessed_on_class: CallableTypeOf[C1.method], |
| 27 | + accessed_on_instance: CallableTypeOf[C1().method], |
| 28 | +): |
| 29 | + reveal_type(accessed_on_class) # revealed: (self: C1, x: int) -> str |
| 30 | + reveal_type(accessed_on_instance) # revealed: (x: int) -> str |
| 31 | +``` |
| 32 | + |
| 33 | +Other callable objects (`staticmethod` objects, instances of classes with a `__call__` method but no |
| 34 | +dedicated `__get__` method) are *not* bound-method descriptors. If accessed as class attributes via |
| 35 | +an instance, they are simply themselves: |
| 36 | + |
| 37 | +```py |
| 38 | +class NonDescriptorCallable2: |
| 39 | + def __call__(self, c2: C2, x: int) -> str: |
| 40 | + return str(x) |
| 41 | + |
| 42 | +class C2: |
| 43 | + non_descriptor_callable: NonDescriptorCallable2 = NonDescriptorCallable2() |
| 44 | + |
| 45 | +def _( |
| 46 | + accessed_on_class: CallableTypeOf[C2.non_descriptor_callable], |
| 47 | + accessed_on_instance: CallableTypeOf[C2().non_descriptor_callable], |
| 48 | +): |
| 49 | + reveal_type(accessed_on_class) # revealed: (c2: C2, x: int) -> str |
| 50 | + reveal_type(accessed_on_instance) # revealed: (c2: C2, x: int) -> str |
| 51 | +``` |
| 52 | + |
| 53 | +Both kinds of objects can inhabit the same `Callable` type: |
| 54 | + |
| 55 | +```py |
| 56 | +class NonDescriptorCallable3: |
| 57 | + def __call__(self, c3: C3, x: int) -> str: |
| 58 | + return str(x) |
| 59 | + |
| 60 | +class C3: |
| 61 | + def method(self: C3, x: int) -> str: |
| 62 | + return str(x) |
| 63 | + |
| 64 | + non_descriptor_callable: NonDescriptorCallable3 = NonDescriptorCallable3() |
| 65 | + |
| 66 | + callable_m: Callable[[C3, int], str] = method |
| 67 | + callable_n: Callable[[C3, int], str] = non_descriptor_callable |
| 68 | +``` |
| 69 | + |
| 70 | +However, when they are accessed on instances of `C3`, they have different signatures: |
| 71 | + |
| 72 | +```py |
| 73 | +def _( |
| 74 | + method_accessed_on_instance: CallableTypeOf[C3().method], |
| 75 | + callable_accessed_on_instance: CallableTypeOf[C3().non_descriptor_callable], |
| 76 | +): |
| 77 | + reveal_type(method_accessed_on_instance) # revealed: (x: int) -> str |
| 78 | + reveal_type(callable_accessed_on_instance) # revealed: (c3: C3, x: int) -> str |
| 79 | +``` |
| 80 | + |
| 81 | +This leaves the question how the `callable_m` and `callable_n` attributes should be treated when |
| 82 | +accessed on instances of `C3`. If we treat `Callable` as being equivalent to a protocol that defines |
| 83 | +a `__call__` method (and no `__get__` method), then they should show no bound-method behavior. This |
| 84 | +is what we currently do: |
| 85 | + |
| 86 | +```py |
| 87 | +reveal_type(C3().callable_m) # revealed: (C3, int, /) -> str |
| 88 | +reveal_type(C3().callable_n) # revealed: (C3, int, /) -> str |
| 89 | +``` |
| 90 | + |
| 91 | +However, this leads to unsoundness: `C3().callable_m` is actually `C3.method` which *is* a |
| 92 | +bound-method descriptor. We currently allow the following call, which will fail at runtime: |
| 93 | + |
| 94 | +```py |
| 95 | +C3().callable_m(C3(), 1) # runtime error! ("takes 2 positional arguments but 3 were given") |
| 96 | +``` |
| 97 | + |
| 98 | +If we were to treat `Callable`s as bound-method descriptors, then the signatures of `callable_m` and |
| 99 | +`callable_n` when accessed on instances would bind the `self` argument: |
| 100 | + |
| 101 | +- `C3().callable_m`: `(x: int) -> str` |
| 102 | +- `C3().callable_n`: `(x: int) -> str` |
| 103 | + |
| 104 | +This would be equally unsound, because now we would allow a call to `C3().callable_n(1)` which would |
| 105 | +also fail at runtime. |
| 106 | + |
| 107 | +There is no perfect solution here, but we can use some heuristics to improve the situation for |
| 108 | +certain use cases (at the cost of purity and simplicity). |
| 109 | + |
| 110 | +## Use case: Decorating a method with a `Callable`-typed decorator |
| 111 | + |
| 112 | +A commonly used pattern in the ecosystem is to use a `Callable`-typed decorator on a method with the |
| 113 | +intention that it shouldn't influence the method's descriptor behavior. For example: |
| 114 | + |
| 115 | +```py |
| 116 | +from typing import Callable |
| 117 | + |
| 118 | +# TODO: this could use a generic signature, but we don't support |
| 119 | +# `ParamSpec` and solving of typevars inside `Callable` types. |
| 120 | +def memoize(f: Callable[[C, int], str]) -> Callable[[C, int], str]: |
| 121 | + raise NotImplementedError |
| 122 | + |
| 123 | +class C: |
| 124 | + def method(self, x: int) -> str: |
| 125 | + return str(x) |
| 126 | + |
| 127 | + @memoize |
| 128 | + def method_decorated(self, x: int) -> str: |
| 129 | + return str(x) |
| 130 | + |
| 131 | +C().method(1) |
| 132 | + |
| 133 | +# TODO: We shouldn't issue any errors here |
| 134 | +# error: [missing-argument] |
| 135 | +# error: [invalid-argument-type] |
| 136 | +C().method_decorated(1) |
| 137 | +``` |
0 commit comments