13 minute read

I was recently surprised to discover that a module called typing had been added to Python 3, providing the fundamental building blocks for constructing types. Although these types are simple hints and have no semantics defined in the language, third party libraries can make use of them to do static type checking.

A project called Mypy, initially developed by a Dropbox team, is able to statically check Python code using a powerful type system with very interesting and useful features, like bidirectional type inference and generics.

If you have some experience with ML languages like Haskell, OCaml or F#, you can probably already see the great value of such a tool; if you don’t, I’m going to point out a few benefits and reference some resources for a better understanding.

Benefits

These are just a few benefits, the ones I observed quickly after starting using Mypy.

Readability

As well mentioned in a Zulip’s blog post, readability is one the biggest and immediate benefits of annotating types for statically type checking. Even when using dynamic languages, it’s a good practice to describe, whenever possible, the signature of a function (How to Design Programs examples), noting the input and output types. This simple practice can save a lot of time when debugging a code, since you wouldn’t need to trace back a chain of function calls to determine the type of some data.

However, if the signature of functions can be only described by comments or docstrings, we have another possible point of failure: buggy signatures. It is very common to change the input/output types of functions over its lifetime, but there is no guarantee that their documentation are consistent if they are not automatically checked.

For example, let’s consider the function argument s3path being a string like s3://my-bucket/path/to/key:

from typing import List

def split_bucket_key(s3path: str) -> List[str]:
    return s3path.split('//')[1].split('/', maxsplit=1)

This code successfully passes the type checking. But then you decide that returning a tuple makes a better interface and you accidentally forget to update the type annotation:

from typing import List

def split_bucket_key(s3path: str) -> List[str]:
    return tuple(
        s3path.split('//')[1].split('/', maxsplit=1)
    )

Mypy would correctly complain about your types:

pycode.py:4: error: Incompatible return value type
(got "Tuple[str, ...]", expected "List[str]")

Fewer Unit Tests

Given enough type annotations, unit tests composed of runtime type checking could be just replaced by the general static type checking done by Mypy. The type inference engine could spot bugs in many different places, specially as new code is added to the code base.

The examples in the next subsections illustrate this idea.

Debugging

In addition to readability, describing function signatures helps debugging, specially if you can trust this information. Instead of worrying about types and values, we can focus only on values, and this can be a tremendous difference.

In the example below we have two versions of an hypothetical function whose purpose is to make a POST request. The first version relies entirely on docstrings as its documentation, while the second uses type annotation that can be checked using Mypy.

import random
from typing import Optional

## Version 1
def post_some_data(url, data):
    """Makes a POST request to the given "url" using "data"
    as a payload and returns the response's body as a string.
    If some error occurs (4xx or 5xx response), "None" is
    returned.
    * url: string
    * data: dictionary
    """
    # Make the real request...
    return random.choice(['something useful', None])

## Version 2
def post_some_data(url: str, data: dict) -> Optional[str]:
    """Returns the response's body for successful requests;
    otherwise, returns "None".
    """
    # Make the real request...
    return random.choice(['something useful', None])

Suppose somewhere in our code base we use this function, but we forget that it not always returns a string:

resp_err = post_some_data('www.blabla.com', {'key': 'value'})
# An accidental mistake.
resp_err.split(' ')

Using the first version, we could only notice this bug when a request fails, at runtime (or we can never get it at all):

Traceback (most recent call last):
  File "pycode.py", line 9, in <module>
    a.split(' ')
AttributeError: 'NoneType' object has no attribute 'split'

If we write the same buggy code using the second version and call Mypy, we instantly get the bug without even executing the code:

$ mypy --strict-optional pycode.py

pycode.py:8: error: Item "None" of "Optional[str]" has no
attribute "split"

Gradual Typing

This seems to be a kind of official term to represent type systems in which dynamic and static types are used together. Some parts could be type checked at compile time (or non-runtime in Python) and others only at runtime. Mypy supports this mixed approach; it keeps the freedom to those who likes the Python dynamic nature to quickly prototype ideas while providing tools to confirm the correctness (in terms of types) of critical parts of an application.

As an example, suppose your application stores some kind of request into a database and each one has its own state. Usually we want to guarantee that nobody accidentally insert invalid values and we end up with a code similar to this one:

def dbexec(data0, data1):
    "A function from a database library."
    pass

STATE_0 = 'posted'
STATE_1 = 'processing'
STATE_2 = 'finished'

def update_state(request, state):
    if state not in set([STATE_0, STATE_1, STATE_2]):
        raise(Exception("Invalid request state"))
    dbexec(request['id'], state)

If a new state constant is added, the set in update_state also needs to be updated. We can improve this code and make it safer by using an Enum class to enumerate the valid request states:

from enum import Enum

def dbexec(data0, data1):
    "A function from a database library."
    pass

class ReqState(Enum):
    STATE_0 = 'posted'
    STATE_1 = 'processing'
    STATE_2 = 'finished'

def update_state(request, state):
    if not isinstance(state, ReqState):
        raise(Exception("Invalid request state"))
    dbexec(request['id'], state.value)

Although it’s better than before, we still can only rely on runtime checking and unit tests. Mypy give us the possibility of checking the code before running it. Using type annotations, the code would be like this:

from enum import Enum

def dbexec(data0, data1):
    "A function from a database library."
    pass

class ReqState(Enum):
    STATE_0 = 'posted'
    STATE_1 = 'processing'
    STATE_2 = 'finished'


def update_state(request: dict, state: ReqState):
    ## Running Mypy before starting the application makes
    ## this checking unnecessary.
    # if not isinstance(state, ReqState):
    #        raise(Exception("Invalid request state"))
    dbexec(request['id'], state.value)

Now the state validation is done before runtime, making the previous isinstance statement unnecessary and the code base safer against many common, accidental bugs. Of course, if the state is expected to come from some IO operation, the runtime checking can’t be avoided.

Another interesting detail is the type definition of the argument request. We can be as much specific as we want about a type. It says that a request must be a dictionary, but it could also say that it shall be a dictionary like Dict[str, int] (string keys, integer values) or even define a very specific structure using a dict like class. Or we could even don’t specify any type at all and still have state being statically checked (as it is done with the return type).

Usage

Annotation syntax

Mypy syntax cheat sheet is a really good reference, all the most used stuff in a single page.

Running Mypy

After installing Mypy, you just need to run it in your project directory (flags explanation):

mypy --ignore-missing-imports --strict-optional project-dir/

Mypy can also be used as a linter in Vim using Syntastic.

Add it to your CI

This is one of the most useful ways to use Mypy: a test stage in your CI pipeline. I use to add it as a step just before the tests execution; if the type checking fails, I don’t even bother to run the tests and the pipeline is marked as failed.

Conclusion

The inclusion of type annotation syntax and the development of a static type checker such as Mypy are a great step the Python community has taken towards safer and clearer code, without giving up the language’s dynamic nature.

References