python dataclasses

Topics

• clean code

Python stdlib’s dataclasses are lightweight and really nice to have. They provide many things out of the box, such as asdict, astuple, __repr__ and allow comparison between objects as well if you pass @dataclass(..., order=True). They should be used if you have few methods and need to remove the boilerplate when defining classes (__init__, __eq__, etc). For more sophisticated usecases where we need to data-validation, use Pydantic dataclasses.

Warning

The frozen=True makes the dataclass immutable, but the attributes might not be immutable (e.g. you can still do: obj.attr3.append(4)). Earlier people used NamedTuple to create such immutable objects.

For immutable objects (created using frozen=True), we would likely want to modify attrs at some point. Take an example of a Point object and a function translate whose job is to move that point by some deltas:

@dataclass(frozen=True)
class Point:
    x: float
    y: float
    color: str # color of point: blue, green etc
 
def translate(point: Point, delta_x: float, delta_y: float) -> Point:
    # we can't do point.x += delta_x since it's frozen/immutable
    return Point(
        x = point.x + delta_x,
        y = point.y + delta_y,
        color = point.color,
    )

Above implementation works but not the best, since we also need to copy attrs which we aren’t even dealing with in the translation process: color = point.color. A better way is to use dataclasses.replace(...) built-in helper func that allows us to update attrs of a dataclass and return a new object. We don’t need to pass attrs which we aren’t changing:

def translate(point: Point, delta_x: float, delta_y: float) -> Point:
    # we can't do point.x += delta_x since it's frozen/immutable
    return dataclasses.repalce(
        point,
        x = point.x + delta_x,
        y = point.y + delta_y,
        # color = point.color is not needed here
    )
 
point = Point(3, 4, "red")
new_point = translate(point, 1, 2)

Related