Variance

Variance describes how subtyping works with generics. Let's take a look at a few examples to set the stage.

Example 1. Iterables

Suppose that we have this class hierarchy¹:

class Animal:
    def __init__(self, name: str) -> None:
        self._name = name

    def name(self) -> str:
        return self._name


class Dog(Animal):
    def bark(self) -> None:
        print("Woof!")

And we have this function:

from collections.abc import Iterable


def greet_animals(animals: Iterable[Animal]) -> None:
    for animal in animals:
        print("Hi, {0}!".format(animal.name()))

Should this call be allowed?

dogs: Iterable[Dog] = [Dog("alice"), Dog("bob"), Dog("charlie")]
greet_animals(dogs)

The answer is yes! All we can do with an Iterable is iterate over it. Every element of the iterable will be a Dog, which is a subtype of Animal. Our code can't assume anything else about the elements, so it should work.

So as you can see, Dog is a subtype of Animal, therefore Iterable[Dog] is a subtype of Iterable[Animal]. But that's not true for all generic types.

Example 2. Lists

Let's take a slightly different function:

def greet_animals(animals: list[Animal]) -> None:
    for animal in animals:
        print("Hi, {0}!".format(animal.name()))

Should this call be allowed?

dogs: list[Dog] = [Dog("alice"), Dog("bob"), Dog("charlie")]
greet_animals(dogs)

If we run pyright on this code, this is what we see:

Argument of type "list[Dog]" cannot be assigned to parameter "animals" of type "list[Animal]" in function "greet_animals"

How so? The call will work fine at runtime, so why not allow it?

Well, for all the type checker knows, greet_animals could do this:

class Cat(Animal):
    def meow(self) -> None:
        print("Meow!")


def greet_animals(animals: list[Animal]) -> None:
    animals.append(Cat("dennis"))
    animals[0] = Animal("emily")

    for animal in animals:
        print("Hi, {0}!".format(animal.name()))

The contract of list[Animal] is that you can push any Animal to it. list[Dog] doesn't satisfy this contract: there are Animals you cannot push into a list[Dog], for example a Cat or a plain Animal object.

This call is also not allowed:

def greet_dogs(dogs: list[Dog]) -> None:
    for dog in dogs:
        print("Hi, {0}!".format(dog.name()))
        dog.bark()

animals: list[Animal] = [Animal("alice"), Cat("bob"), Dog("charlie")]
greet_dogs(animals)

A list[Animal] can already contain animals that are not Dogs, so the assumptions of the function (namely, the fact that you can call bark() on each element in the list) will be violated.

In other words, list[Dog] is not a subtype of list[Animal], and list[Animal] is not a subtype of list[Dog] either.

Example 3. Functions

Consider this function:

from collections.abc import Callable

def create_animals(factory: Callable[[str], Animal]) -> tuple[Animal, Animal]:
    return factory("alice"), factory("bob")

Which of these functions would work as the factory?

from typing import Union


class BetterString(str):
    ...

    def to_awesome_case(self) -> str:
        import random
        return "".join(random.choice([char.lower(), char.upper()]) for char in self)


def make_animal(name: str) -> Animal: ...
def make_cat(name: str) -> Cat: ...
def make_cat_or_dog(name: str) -> Union[Cat, Dog]: ...
def make_cat2(name_or_id: Union[str, int]) -> Cat: ...
def make_cat3(whatever: object) -> Cat: ...
def make_animal2(name: BetterString) -> Animal: ...
def make_animal_of_plant(name: str) -> Union[Animal, Plant]: ...

def make_animal(name: str) -> Animal: ...

This function has exactly the type we need - Callable[[str], Animal]. So yes, it will work
def make_cat(name: str) -> Cat: ...

This function returns a Cat instead of just any animal. But that's fine: we need a function that returns an Animal, and whatever make_cat returns is an Animal.

You can think of it in another way: you could make a trivial function that transforms this into (name: str) -> Animal:
```
def make_cat_wrapper(name: str) -> Animal:
    animal: Animal = make_cat(name)
    return animal
```
def make_cat_or_dog(name: str) -> Union[Cat, Dog]: ...

Same for this function: whatever it returns, it is an Animal.
def make_cat2(name_or_id: Union[str, int]) -> Cat: ...

This one is more tricky, but it will still fit: we need a function that accepts a string as an argument, and this function satisfies this requirement.

In other words, make_cat2 accepts any subtype of Union[str, int] as an argument, and str is a subtype of Union[str, int].
def make_cat3(whatever: object) -> Cat: ...

This is also fine. make_cat3 will work with any object at all, including a string.
def make_animal2(name: BetterString) -> Animal: ...

This one will not fit. We need our factory to accept any string, but make_animal2 only accepts BetterString objects and can, for example, call the to_awesome_case() method on the name.
def make_animal_of_plant(name: str) -> Union[Animal, Plant]: ...

This will not work either. The function should always return an Animal.

As you can see, there are two rules for functions:

If Child is a subtype of Parent, then Callable[[X], Child] is a subtype of Callable[[X], Parent]
If Child is a subtype of Parent, then Callable[[Parent], X] is a subtype of Callable[[Child], X]

Definition

Suppose that we have a generic type F[T], and two ordinary types: Child and Parent, where Child is a subtype of Parent. The variance of F describes how F[Child] is related to F[Parent].

There are three kinds of variance:

Covariance: F[Child] is a subtype of F[Parent]
- example: tuple[Dog, ...] is a subtype of tuple[Animal, ...]
- example: Callable[[str], Dog] is a subtype of Callable[[str], Animal]
Contravariance: F[Parent] is a subtype of F[Child]
- example: Callable[[Animal], None] is a subtype of Callable[[Cat], None]
Invariance: F[Child] is not related to F[Parent]
- example: list[Dog] is not related to list[Animal]

When describing a type, replace the "ance" suffix with "ant". For example, list is invariant, while tuple is covariant.

More than one type variable

Some generic types take in more than one type variable. For example, Callable can take any number of variables for parameters, and then another one for the result type.

In that case, a generic type can have different variance in different type variables.

For example:

Callable is contravariant in the parameter types but covariant in the return type
Mapping is invariant in the key type but covariant in the return type

I know, I hate nonsense examples of inheritance as well. I hope that you forgive me. ↩