Python static type checker (mypy)

mypy is a compile-time static type checker for Python, allowing optional, gradual typing of Python code. Zulip was fully annotated with mypy’s Python 2 syntax in 2016, before our migration to Python 3 in late 2017. In 2018 and 2020, we migrated essentially the entire codebase to the nice PEP 484 (Python 3 only) and PEP 526 (Python 3.6) syntax for static types:

user_dict: Dict[str, UserProfile] = {}

def get_user(email: str, realm: Realm) -> UserProfile:
    ... # Actual code of the function here

You can learn more about it at:

The mypy type checker is run automatically as part of Zulip’s Travis CI testing process in the backend build.

Installing mypy

mypy is installed by default in the Zulip development environment. If you’d like to install just the version of mypy that we’re using (useful if e.g. you want mypy installed on your laptop outside the Vagrant guest), you can do that with pip install -r requirements/mypy.txt.

Running mypy on Zulip’s code locally

To run mypy on Zulip’s python code, you can run the command:

tools/run-mypy

Mypy outputs errors in the same style as a compiler would. For example, if your code has a type error like this:

foo = 1
foo = '1'

you’ll get an error like this:

test.py: note: In function "test":
test.py:200: error: Incompatible types in assignment (expression has type "str", variable has type "int")

Mypy is there to find bugs in Zulip before they impact users

For the purposes of Zulip development, you can treat mypy like a much more powerful linter that can catch a wide range of bugs. If, after running tools/run-mypy on your Zulip branch, you get mypy errors, it’s important to get to the bottom of the issue, not just do something quick to silence the warnings, before we merge the changes. Possible explanations include:

  • A bug in any new type annotations you added.

  • A bug in the existing type annotations.

  • A bug in Zulip!

  • Some Zulip code is correct but confusingly reuses variables with different types.

  • A bug in mypy (though this is increasingly rare as mypy is now fairly mature as a project).

Each explanation has its own solution, but in every case the result should be solving the mypy warning in a way that makes the Zulip codebase better. If you’re having trouble, silence the warning with an Any or # type: ignore[code] so you’re not blocked waiting for help, add a # TODO: comment so it doesn’t get forgotten in code review, and ask for help in chat.zulip.org.

Mypy stubs for third-party modules

For the Python standard library and some popular third-party modules, the typeshed project has stubs, basically the equivalent of C header files defining the types used in these Python APIs.

For other third-party modules that we call from Zulip, one either needs to add an ignore_missing_imports entry in mypy.ini in the root of the project, letting mypy know that it’s third-party code, or add type stubs to the stubs/ directory, which has type stubs that mypy can use to type-check calls into that third-party module.

It’s easy to add new stubs! Just read the docs, look at some of existing examples to see how they work, and remember to remove the ignore_missing_imports entry in mypy.ini when you add them.

For any third-party modules that don’t have stubs, mypy treats everything in the third-party module as an Any, which is the right model (one certainly wouldn’t want to need stubs for everything just to use mypy!), but means the code can’t be fully type-checked.

type_debug.py

zerver/lib/type_debug.py has a useful decorator print_types. It prints the types of the parameters of the decorated function and the return type whenever that function is called. This can help find out what parameter types a function is supposed to accept, or if parameters with the wrong types are being passed to a function.

Here is an example using the interactive console:

>>> from zerver.lib.type_debug import print_types
>>>
>>> @print_types
... def func(x, y):
...     return x + y
...
>>> func(1.0, 2)
func(float, int) -> float
3.0
>>> func('a', 'b')
func(str, str) -> str
'ab'
>>> func((1, 2), (3,))
func((int, int), (int,)) -> (int, int, int)
(1, 2, 3)
>>> func([1, 2, 3], [4, 5, 6, 7])
func([int, ...], [int, ...]) -> [int, ...]
[1, 2, 3, 4, 5, 6, 7]

print_all prints the type of the first item of lists. So [int, ...] represents a list whose first element’s type is int. Types of all items are not printed because a list can have many elements, which would make the output too large.

Similarly in dicts, one key’s type and the corresponding value’s type are printed. So {1: 'a', 2: 'b', 3: 'c'} will be printed as {int: str, ...}.

Using @overload to accurately describe variations

Sometimes, a function’s type is most precisely expressed as a few possibilities, and which possibility can be determined by looking at the arguments. You can express that idea in a way mypy understands using @overload. For example, check_list returns a Validator function that verifies that an object is a list, raising an exception if it isn’t.

It supports being passed a sub_validator, which will verify that each element in the list has a given type as well. One can express the idea “If sub_validator validates that something is a ResultT, check_list(sub_validator) validators that something is a List[ResultT] as follows:

@overload
def check_list(sub_validator: None, length: Optional[int]=None) -> Validator[List[object]]:
    ...
@overload
def check_list(sub_validator: Validator[ResultT],
               length: Optional[int]=None) -> Validator[List[ResultT]]:
    ...
def check_list(sub_validator: Optional[Validator[ResultT]]=None,
               length: Optional[int]=None) -> Validator[List[ResultT]]:

The first overload expresses the types for the case where no sub_validator is passed, in which case all we know is that it returns a Validator[List[object]]; whereas the second defines the type logic for the case where we are passed a sub_validator.

Warning: Mypy only checks the body of an overloaded function against the final signature and not against the more restrictive @overload signatures. This allows some type errors to evade detection by mypy:

@overload
def f(x: int) -> int: ...
@overload
def f(x: str) -> int: ...  # oops
def f(x: Union[int, str]) -> Union[int, str]:
    return x

x: int = f("three!!")

Due to this potential for unsafety, we discourage overloading unless it’s absolutely necessary. Consider writing multiple functions with different names instead.

See the mypy overloading documentation for more details.

Best practices

When is a type annotation justified?

Usually in fully typed code, mypy will protect you from writing a type annotation that isn’t justified by the surrounding code. But when you need to write annotations at the border between untyped and typed code, keep in mind that a type annotation should always represent a guarantee, not an aspiration. If you have validated that some value is an int, it can go in an int annotated variable. If you are going to validate it later, it should not. When in doubt, an object annotation is always safe.

Mypy understands many Python constructs like assert, if, isinstance, and logical operators, and uses them to automatically narrow the type of validated objects in many cases.

def f(x: object, y: Optional[str]) -> None:
    if isinstance(x, int):
        # Within this if block, mypy infers that x: int
        print(x + 1)
    assert y is not None
    # After that assert statement, mypy infers that y: str
    print(y.strip())

It won’t be able do this narrowing if the validation is hidden behind a function call, so sometimes it’s helpful for a validation function to return the type-narrowed value back to the caller even though the caller already has it. (The validators in zerver/lib/validator.py are examples of this pattern.)

Avoid the Any type

Mypy provides the Any type for interoperability with untyped code, but it is completely unchecked. You can put an value of an arbitrary type into an expression of type Any, and get an value of an arbitrary type out, and mypy will make no effort to check that the input and output types match. So using Any defeats the type safety that mypy would otherwise provide.

x: Any = 5
y: str = x  # oops
print(y.strip())  # runtime error

If you think you need to use Any, consider the following safer alternatives first:

  • To annotate a dictionary where different keys correspond to values of different types, instead of writing Dict[str, Any], try declaring a dataclass or a TypedDict.

  • If you’re annotating a class or function that might be used with different data types at different call sites, similar to the builtin List type or the sorted function, generic types with TypeVar might be what you need.

  • If you need to accept data of several specific possible types at a single site, you may want a Union type. Union is checked: before using value: Union[str, int] as a str, mypy requires that you validate it with an instance(value, str) test.

  • If you really have no information about the type of a value, use the object type. Since every type is a subtype of object, you can correctly annotate any value as object. The difference between Any and object is that mypy will check that you safely validate an object with isinstance before using it in a way that expects a more specific type.

  • A common way for Any annotations to sneak into your code is the interaction with untyped third-party libraries. Mypy treats any value imported from an untyped library as annotated with Any, and treats any type imported from an untyped library as equivalent to Any. Consider providing real type annotations for the library by writing a stub file.

Avoid cast()

The cast function lets you provide an annotation that Mypy will not verify. Obviously, this is completely unsafe in general.

x = cast(str, 5)  # oops
print(x.strip())  # runtime error

Instead of using cast:

  • You can use a variable annotation to be explicit or to disambiguate types that mypy can check but cannot infer.

    l: List[int] = []
    
  • You can use an isinstance test to safely verify that a value really has the type you expect.

Avoid # type: ignore comments

Mypy allows you to ignore any type checking error with a # type: ignore comment, but you should avoid this in the absence of a very good reason, such as a bug in mypy itself. If there are no safe options for dealing with the error, prefer an unchecked cast, since its unsafety is somewhat more localized.

Our linter requires all # type: ignore comments to be scoped to the specific error code being ignored, and followed by an explanation such as a link to a GitHub issue.

Avoid other unchecked constructs

  • As mentioned above, we discourage writing overloaded functions because their bodies are not checked against the @overload signatures.

  • Avoid Callable[..., T] (with literal ellipsis ...), since mypy cannot check the types of arguments passed to it. Provide the specific argument types (Callable[[int, str], T]) in simple cases, or use callback protocols in more complex cases.

Use Optional and None correctly

The Optional type is for optional values, which are values that could be None. For example, Optional[int] is equivalent to Union[int, None].

The Optional type is not for optional parameters (unless they are also optional values as above). This signature does not use the Optional type:

def func(flag: bool = False) -> str:
    ...

A collection such as List should only be Optional if None would have a different meaning than the natural meaning of an empty collection. For example:

  • An include list where the default is to include everything should be Optional with default None.

  • An exclude list where the default is to exclude nothing should be non-Optional with default [].

Don’t test an Optional value using truthiness (if value:, not value, value or default_value), especially when the type might have falsy values other than None.

s: Optional[str]
if not s:  # bad: are we checking for None or ""?
    ...
if s is None:  # good
    ...

Read-only types

The basic Python collections List, Dict, and Set are mutable, but it’s confusing for a function to mutate a collection that was passed to it as an argument, especially by accident. To avoid this, prefer annotating function parameters with read-only types:

  • Sequence or Iterable instead of List,

  • Mapping instead of Dict,

  • AbstractSet instead of Set.

This is especially important for parameters with default arguments, since a mutable default argument is confusingly shared between all calls to the function.

def f(items: Sequence[int] = []) -> int:
    items.append(1)  # mypy catches this mistake
    return sum(items)

A function’s return type can be mutable if the return value is always a freshly created collection, since the caller ends up with the only reference to the value and can freely mutate it without risk of confusion. But a read-only return type might be more appropriate for a function that returns a reference to an existing collection.

Read-only types have the additional advantage of being covariant rather than invariant: if B is a subtype of A, then List[B] may not be converted to List[A], but Sequence[B] may be converted to Sequence[A].

Typing decorators

A simple decorator that operates on functions of a fixed signature works with no issues:

def fancy(func: Callable[[str], str]) -> Callable[[int], str]:
    def wrapped_func(n: int) -> str:
        print("so fancy")
        return func(str(n))
    return wrapped_func

@fancy
def f(s: str) -> str:
    return s

A decorator with an argument also works:

def fancy(message: str) -> Callable[[Callable[[str], str]], Callable[[int], str]]:
    def wrapper(func: Callable[[str], str]) -> Callable[[int], str]:
        def wrapped_func(n: int) -> str:
            print(message)
            return func(str(n))
        return wrapped_func
    return wrapper

@fancy("so fancy")
def f(s: str) -> str:
    return s

And a generic decorator that operates on functions of arbitrary signatures can be written with a cast if the output signature is always the same as the input signature:

FuncT = TypeVar("FuncT", bound=Callable[..., object])

def fancy(func: FuncT) -> FuncT:
    def wrapped_func(*args: object, **kwargs: object) -> object:
        print("so fancy")
        return func(*args, **kwargs)
    return cast(FuncT, wrapped_func)  # https://github.com/python/mypy/issues/1927

@fancy
def f(s: str) -> str:
    return s

(A generic decorator with an argument would return Callable[[FuncT], FuncT].)

But Mypy doesn’t yet support the advanced type annotations that would be needed to correctly type generic signature-changing decorators, such as zerver.decorator.authenticated_json_view, which passes an extra argument to the inner function. For these decorators we must unfortunately give up some type safety by falling back to Callable[..., T].

Troubleshooting advice

All of our linters, including mypy, are designed to only check files that have been added in git (this is by design, since it means you have untracked files in your Zulip checkout safely). So if you get a mypy error like this after adding a new file that is referenced by the existing codebase:

mypy | zerver/models.py:1234: note: Import of 'zerver.lib.markdown_wrappers' ignored
mypy | zerver/models.py:1234: note: (Using --follow-imports=error, module not passed on command line)

The problem is that you need to git add the new file.