IO is ugly.

Why? Let me illustrate it with the example.

IO marker

Imagine we have this beautiful pure function:

def can_book_seats(
    number_of_seats: int,
    reservation: 'Reservation',
) -> bool:
    return reservation.capacity >= number_of_seats + reservation.booked

What’s good about it? We can test it easily. Even without setting up any testing framework, simple doctests will be enough.

This code is beautiful, because it is simple.

We can later use its result to notify users about their booking request:

def notify_user_about_booking_result(is_successful: bool) -> 'MessageID':

notify_user_about_booking_result(is_successful)  # works just fine!

Impure functions

But, imagine that our requirements had changed. And now we have to grab the number of already booked tickets from some other provider and fetch the maximum capacity from the database:

import requests
import db

def can_book_seats(
    number_of_seats: int,
    place_id: int,
) -> bool:
    capacity = db.get_place_capacity(place_id)  # sql query
    booked = requests('').json()['booked']  # http req
    return capacity >= number_of_seats + booked

Now testing this code will become a nightmare! It will require to setup:

  • real database and tables

  • fixture data

  • requests mocks for different outcomes

  • and the whole Universe!

Our complexity has sky-rocketed! And the most annoying part is that all other functions that call can_book_seats now also have to do the same setup. It seams like IO is indelible mark (some people also call it “effect”).

And at some point it time we will start to mix pure and impure code together.

Separating two worlds

Well, our IO mark is indeed indelible and should be respected.

Once you have an IO operation you can mark it appropriately. And it infects all other functions that call it. And impurity becomes explicit:

import requests
import db
from import IO

def can_book_seats(
    number_of_seats: int,
    place_id: int,
) -> IO[bool]:
    capacity = db.get_place_capacity(place_id)  # sql query
    booked = requests('').json()['booked']
    return IO(capacity >= number_of_seats + booked)

Now this function returns IO[bool] instead of a regular bool. It means, that it cannot be used where regular bool can be:

def notify_user_about_booking_result(is_successful: bool) -> 'MessageID':

is_successful: IO[bool] = can_book_seats(number_of_seats, place_id)
notify_user_about_booking_result(is_successful)  # Boom!
# => Argument 1 has incompatible type "IO[bool]"; expected "bool"

See? It is now impossible for a pure function to use IO[bool]. It is impossible to unwrap or get a value from this container. Once it is marked as IO it will never return to the pure state (well, there’s a hack actually: unsafe_perform_io).

IO container also needs to be explicitly mapped to produce new IO result:

message_id: IO['MessageID'] = can_book_seats(

Or it can be annotated to work with impure results:

def notify_user_about_booking_result(
    is_successful: IO[bool],
) -> IO['MessageID']:

is_successful: IO[bool] = can_book_seats(number_of_seats, place_id)
notify_user_about_booking_result(is_successful)  # Works!

Now, all our impurity is explicit. We can track it, we can fight it, we can design it better. By saying that, it is assumed that you have a functional core and imperative shell.


You can also lift regular function into one that works with IO on both ends. It really helps you with the composition!

def regular_function(arg: int) -> float:
    return arg / 2  # not an `IO` operation

container: IO[int]
# When we need to compose `regular_function` with `IO`,
# we have two ways of doing it:
# or, it is the same as:

The second variant is useful when using returns.pipeline.pipe() and other different declarative tools.


We also have this handy decorator to help you with the existing impure things in Python:

from import impure

name: IO[str] = impure(input)('What is your name?')

You can also decorate your own functions with @impure for better readability and clearness:

import requests
from import impure

def get_user() -> 'User':
    return requests.get('https:...').json()


Typing will only work correctly if decorator_plugin is used. This happens due to mypy issue.


Sometimes you really need to get the raw value from IO container. For example:

def index_view(request, user_id):
    user: IO[User] = get_user(user_id)
    return render('index.html', { user: user })  # ???

In this case your web-framework will not render your user correctly. Since it does not expect it to be wrapped inside IO containers. And we obviously cannot map or bind this function.

What to do? Use unsafe_perform_io:

from returns.unsafe import unsafe_perform_io

def index_view(request, user_id):
    user: IO[User] = get_user(user_id)
    return render('index.html', { user: unsafe_perform_io(user) })  # Ok

We need it as an escape and compatibility mechanism for our imperative shell.

It is recommended to use import-linter to restrict imports from returns.unsafe expect the top-level modules.

Inspired by Haskell’s unsafePerformIO


What is the difference between IO[T] and T?

What kind of input parameter should my function accept IO[T] or simple T?

It really depends on your domain / context. If the value is pure, than use raw unwrapped values. If the value is fetched, input, received, selected, than use IO container.

Most web applications are just covered with IO.

Why IO should be at the top level of composition?

As we state in Composition docs we allow to compose different containers together.

We prefer IO[Result[A, B]] and sticking to the single version allows better composition. The same rule is applied to Maybe and all other containers we have.

Composing IO at the top level is easier because you can join things easily.

And other containers not always make sense. If some operation performs IO it should mark all internals.

Why can’t we unwrap values or use @pipeline with IO?

Our design decision was not let people unwrap IO containers, so it will indeed infect the whole call-stack with its effect.

Otherwise, people might hack the system in some dirty (from our point of view) but valid (from the python’s point of view) ways.


Of course, you can directly access
the internal state of the IO with `._internal_state`,
but your are considered to be a grown-up!

Use wemake-python-styleguide to restrict `._` access in your code.

API Reference

graph TD; IO Generic --> IO BaseContainer --> IO
class IO(inner_value)[source]

Bases: typing.Generic, returns.primitives.container.BaseContainer

Explicit marker for impure function results.

We call it “marker” since once it is marked, it cannot be unmarked.

IO is also a container. But, it is different in a way that it cannot be unwrapped / rescued / fixed. There’s no way to directly get its internal value.


Applies function to the inner value.

Applies ‘function’ to the contents of the IO instance and returns a new IO object containing the result. ‘function’ should accept a single “normal” (non-container) argument and return a non-container result.

>>> def mappable(string: str) -> str:
...      return string + 'b'
>>> IO('a').map(mappable) == IO('ab')
Return type



Applies ‘function’ to the result of a previous calculation.

‘function’ should accept a single “normal” (non-container) argument and return IO type object.

>>> def bindable(string: str) -> IO[str]:
...      return IO(string + 'b')
>>> IO('a').bind(bindable) == IO('ab')
Return type


classmethod lift(function)[source]

Lifts function to be wrapped in IO for better composition.

In other words, it modifies the function’s signature from: a -> b to: IO[a] -> IO[b]

This is how it should be used:

>>> def example(argument: int) -> float:
...     return argument / 2  # not exactly IO action!
>>> IO.lift(example)(IO(2)) == IO(1.0)
Return type

Callable[[IO[+_ValueType]], IO[~_NewValueType]]


Decorator to mark function that it returns IO container.

Supports both async and regular functions.


Compatibility utility and escape mechanism from IO world.

Just unwraps the internal value from IO container. Should be used with caution! Since it might be overused by lazy and ignorant developers.

It is recommended to have only one place (module / file) in your program where you allow unsafe operations.

We recommend to use import-linter to enforce this rule:

>>> from import IO
>>> unsafe_perform_io(IO(1))
Return type


Maybe Result