JelleZijlstra · Jul 19, 2024 · Jul 14, 2024 · Jul 14, 2024 · Jul 14, 2024 · Jul 14, 2024
diff --git a/docs/source/guides.rst b/docs/source/guides.rst
@@ -10,4 +10,5 @@ Type System Guides
    writing_stubs
    modernizing
    unreachable
+   type_narrowing
    typing_anti_pitch
diff --git a/docs/source/type_narrowing.rst b/docs/source/type_narrowing.rst
@@ -0,0 +1,324 @@
+**************
+Type Narrowing
+**************
+
+Python programs often contain values that can take on multiple types and
+that are distinguished by a conditional check at runtime. For example, here
+the variable *name* can be either a ``str`` or ``None``, and the
+``if name is not None`` narrows it down to just ``str``::
+
+    def maybe_greet(name: str | None) -> None:
+        if name is not None:
+            print("Hello, " + name)
+
+This technique is called *type narrowing*.
+In order to avoid false positives on such code, type checkers understand
+various kinds of conditionals that are used to narrow types in Python code.
+The exact set of type narrowing constructs that a type checker understands
+is not specified and varies across type checkers. Commonly understood type
+patterns include:
+
+* ``if x is not None``
+* ``if x``
+* ``if isinstance(x, SomeType)``
+* ``if callable(x)``
+
+In addition to narrowing local variables, type checkers usually also support
+narrowing instance attributes and sequence members, such as
+``if x.some_attribute is not None`` or ``if x[0] is not None``.
+
+Consult your type checker's documentation for more information on the type
+narrowing constructs it supports.
+
+The type system also includes two ways to create *user-defined* type narrowing
+functions: :py:data:`typing.TypeIs` and :py:data:`typing.TypeGuard`. These
+are useful if you want to reuse a more complicated check in multiple places, or
+you use a check that the type checker doesn't understand. In these cases, you
+can define a ``TypeIs`` or ``TypeGuard`` function to perform the check and allow type checkers
+to use it to narrow the type of a variable. Among the two, ``TypeIs`` usually
-to use it to narrow the type of a variable. Among the two, ``TypeIs`` usually
+to use it to narrow the type of a variable. Between the two, ``TypeIs`` usually
-to use it to narrow the type of a variable. Among the two, ``TypeIs`` usually
+to use it to narrow the type of a variable. Between the two, ``TypeIs`` usually
+has the more intuitive behavior, so we'll talk about it more; see
+:ref:`below <guide-type-narrowing-typeis-typeguard>` for a comparison.
+
+How to use ``TypeIs`` and ``TypeGuard``
+---------------------------------------
+
+A ``TypeIs`` function takes a single argument and is annotated as returning
+``TypeIs[T]``, where ``T`` is the type that you want to narrow to. The function
+must return ``True`` if the argument is of type ``T``, and ``False`` otherwise.
+The function can then be used in ``if`` checks, just like you would use ``isinstance()``.
+For example::
+
+    from typing import Literal, TypeIs
+
+    type Direction = Literal["N", "E", "S", "W"]
+
+    def is_direction(x: str) -> TypeIs[Direction]:
+        return x in {"N", "E", "S", "W"}
+
+    def maybe_direction(x: str) -> None:
+        if is_direction(x):
+            print(f"{x} is a cardinal direction")
+        else:
+            print(f"{x} is not a cardinal direction")
+
+A ``TypeGuard`` function looks similar and is used in the same way, but the
+type narrowing behavior is different, as dicussed in :ref:`the section below <guide-type-narrowing-typeis-typeguard>`.
+
+Depending on the version of Python you are running, you will be able to
+import ``TypeIs`` and ``TypeGuard`` either from the standard library :py:mod:`typing`
+module or from the third-party ``typing_extension`` module:
+
+* ``TypeIs`` is in ``typing`` starting from Python 3.13 and in ``typing_extension``
+  starting from version 4.10.0.
+* ``TypeGuard`` is in ``typing`` starting from Python 3.10 and in ``typing_extension``
+  starting from version 3.10.0.0.
+
+
+Writing a correct ``TypeIs`` function
+-------------------------------------
+
+A ``TypeIs`` function allows you to override your type checker's type narrowing
+behavior. This is a powerful tool, but it can be dangerous because an incorrectly
+written ``TypeIs`` function can lead to unsound type checking, and type checkers
+cannot detect such errors.
+
+For a function returning ``TypeIs[T]`` to be correct, it must return ``True`` if and only if
+the argument is compatible with type ``T``, and ``False`` otherwise. If this condition is
+not met, the type checker may infer incorrect types.
+
+Below are some examples of correct and incorrect ``TypeIs`` functions::
+
+    from typing import TypeIs
+
+    # Correct
+    def good_typeis(x: object) -> TypeIs[int]:
+        return isinstance(x, int)
+
+    # Incorrect: does not return True for all ints
+    def bad_typeis1(x: object) -> TypeIs[int]:
+        return isinstance(x, int) and x > 0
+
+    # Incorrect: returns True for some non-ints
+    def bad_typeis2(x: object) -> TypeIs[int]:
+        return isinstance(x, (int, float))
+
+This function demonstrates some errors that can occur when using a poorly written
+``TypeIs`` function. These errors are not detected by type checkers::
+
+    def caller(x: int | str, y: int | float) -> None:
+        if bad_typeis1(x):  # narrowed to int
+            print(x + 1)
+        else:  # narrowed to str (incorrectly)
+            print("Hello " + x)  # runtime error if x is a negative int
+
+        if bad_typeis2(y):  # narrowed to int
+            # Because of the incorrect TypeIs, this branch is taken at runtime if
+            # y is a float.
+            print(y.bit_count())  # runtime error: this method exists only on int, not float
+        else:  # narrowed to float (though never executed at runtime)
+            pass
+
+Here is an example of a correct ``TypeIs`` function for a more complicated type::
+
+    from typing import TypedDict, TypeIs
+
+    class Point(TypedDict):
+        x: int
+        y: int
+
+    def is_point(x: object) -> TypeIs[Point]:
+        return (
+            isinstance(x, dict)
+            and all(isinstance(key, str) for key in x)
+            and "x" in x
+            and "y" in x
+            and isinstance(x["x"], int)
+            and isinstance(x["y"], int)
+        )
+
+.. _`guide-type-narrowing-typeis-typeguard`:
+
+``TypeIs`` and ``TypeGuard``
+----------------------------
+
+:py:data:`typing.TypeIs` and :py:data:`typing.TypeGuard` are both tools for narrowing the type of a variable
+based on a user-defined function. Both can be used to annotate functions that take an
+argument and return a boolean depending on whether the input argument is compatible with
+the narrowed type. These function can then be used in ``if`` checks to narrow the type
+of a variable.
+
+``TypeIs`` usually has the most intuitive behavior, but it
+introduces more restrictions. ``TypeGuard`` is the right tool to use if:
+
+* You want to narrow to a type that is not compatible with the input type, for example
+  from ``list[object]`` to ``list[int]``.  ``TypeIs`` only allows narrowing between
+  compatible types.
+* Your function does not return ``True`` for all input values that are compatible with
+  the narrowed type. For example, you could have a ``TypeGuard[int]`` that returns ``True``
+  only for positive integers.
+
+``TypeIs`` and ``TypeGuard`` differ in the following ways:
+
+* ``TypeIs`` requires the narrowed type to be a subtype of the input type, while
+  ``TypeGuard`` does not.
+* When a ``TypeGuard`` function returns ``True``, type checkers narrow the type of the
+  variable to exactly the ``TypeGuard`` type. When a ``TypeIs`` function returns ``True``,
+  type checkers can infer a more precise type combining the previously known type of the
+  variable with the ``TypeIs`` type. (Technically, this is known as an intersection type.)
+* When a ``TypeGuard`` function returns ``False``, type checkers cannot narrow the type of
+  the variable at all. When a ``TypeIs`` function returns ``False``, type checkers can narrow
+  the type of the variable to exclude the ``TypeIs`` type.
+
+This behavior can be seen in the following example::
+
+    from typing import TypeGuard, TypeIs, reveal_type, final
+
+    class Base: ...
+    class Child(Base): ...
+    @final
+    class Unrelated: ...
+
+    def is_base_typeguard(x: object) -> TypeGuard[Base]:
+        return isinstance(x, Base)
+
+    def is_base_typeis(x: object) -> TypeIs[Base]:
+        return isinstance(x, Base)
+
+    def use_typeguard(x: Child | Unrelated) -> None:
+        if is_base_typeguard(x):
+            reveal_type(x)  # Base
+        else:
+            reveal_type(x)  # Child | Unrelated
+
+    def use_typeis(x: Child | Unrelated) -> None:
+        if is_base_typeis(x):
+            reveal_type(x)  # Child
+        else:
+            reveal_type(x)  # Unrelated
+
+
+Safety and soundness
+--------------------
+
+While type narrowing is important for typing real-world Python code, many
+forms of type narrowing are unsafe in the presence of mutability. Type checkers
+attempt to limit type narrowing in a way that minimizes unsafety while remaining
+useful, but not all safety violations can be detected.
+
+Incorrect ``TypeIs`` and ``TypeGuard`` functions
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Both ``TypeIs`` and ``TypeGuard`` rely on the user writing a function that
+returns whether an object is of a particular type. However, the type checker
+does not validate whether the function actually behaves as expected. If it
+does not, the type checker's narrowing behavior will not match what happens
+at runtime.::
+
+    from typing import TypeIs
+
+    def is_str(x: object) -> TypeIs[str]:
+        return True
+
+    def takes_str_or_int(x: str | int) -> None:
+        if is_str(x):
+            print(x + " is a string")  # runtime error
+
+To avoid this problem, every ``TypeIs`` and ``TypeGuard`` function should be
+carefully reviewed and tested.
+
+Unsound ``TypeGuard`` narrowing
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Unlike ``TypeIs``, ``TypeGuard`` can narrow to a type that is not a subtype of the
+original type. This allows for unsafe behavior with invariant data structures::
+
+    from typing import Any, TypeGuard
+
+    def is_int_list(x: list[Any]) -> TypeGuard[list[int]]:
+        return all(isinstance(i, int) for i in x)
+
+    def maybe_mutate_list(x: list[Any]) -> None:
+        if is_int_list(x):
+            x.append(0)  # OK, x is narrowed to list[int]
+
+    def takes_bool_list(x: list[bool]) -> None:
+        maybe_mutate_list(x)
+        reveal_type(x)  # list[bool]
+        assert all(isinstance(i, bool) for i in x)  # fails at runtime
+
+    takes_bool_list([True, False])
+
+To avoid this problem, use ``TypeIs`` instead of ``TypeGuard`` where possible.
+If you must use ``TypeGuard``, avoid narrowing across incompatible types.
+
+Invalidated assumptions
+~~~~~~~~~~~~~~~~~~~~~~~
+
+One category of safety issues relates to the fact that type narrowing relies
+on a condition that was established at one point in the code and is then relied
+on later: we first check ``if x is not None``, then rely on ``x`` not being ``None``.
+However, in the meantime other code may have run (for example, in another thread,
+another coroutine, or simply some code that was invoked by a function call) and
+invalidated the earlier condition.
+
+Such problems are most likely when narrowing is performed on elements of mutable
+objects, but it is possible to construct unsafe examples even using only narrowing
+of local variables::
+
+    def maybe_greet(name: str | None) -> None:
+        def set_it_to_none():
+            nonlocal name
+            name = None
+
+        if name is not None:
+            set_it_to_none()
+            # fails at runtime, no error in current type checkers
+            print("Hello " + name)
+
+    maybe_greet("Guido")
+
+A more realistic example might involve multiple coroutines mutating a list::
+
+    import asyncio
+    from typing import Sequence, TypeIs
+
+    def is_int_sequence(x: Sequence[object]) -> TypeIs[Sequence[int]]:
+        return all(isinstance(i, int) for i in x)
+
+    async def takes_seq(x: Sequence[int | None]):
+        if is_int_sequence(x):
+            await asyncio.sleep(2)
+            print("The total is", sum(x))  # fails at runtime
+
+    async def takes_list(x: list[int | None]):
+        t = asyncio.create_task(takes_seq(x))
+        await asyncio.sleep(1)
+        x.append(None)
+        await t
+
+    if __name__ == "__main__":
+        lst: list[int | None] = [1, 2, 3]
+        asyncio.run(takes_list(lst))
+
+These issues unfortunately cannot be fully detected by the current
+Python type system. (An example of a different programming language that
+does solve this problem is Rust, which uses a system called
+`ownership <https://doc.rust-lang.org/book/ch04-01-what-is-ownership.html>`__.)
+To avoid such issues, avoid using type narrowing on objects that are mutated
+from other parts of the code.
+
+
+See also
+--------
+
+* Type checker documentation on type narrowing
+
+  * `Mypy <https://mypy.readthedocs.io/en/stable/type_narrowing.html>`__
+  * `Pyright <https://microsoft.github.io/pyright/#/type-concepts-advanced?id=type-narrowing>`__
+
+* PEPs related to type narrowing. These contain additional discussion
+  and motivation for current type checker behaviors.
+
+  * :pep:`647` (introduced ``TypeGuard``)
+  * (*withdrawn*) :pep:`724` (proposed change to ``TypeGuard`` behavior)
+  * :pep:`742` (introduced ``TypeIs``)