Dataclasses and mutable defaults

Posted on Tue 14 January 2020 in programming

One common Python gotcha is the use of mutable objects as defaults for function keyword arguments. There are approximately one billion questions on SO about this or nice discussions elsewhere. I came across a nice feature in Python's dataclasses library that addresses a similar problem.

Mutable defaults are bad

As a reminder, in this example, every invocation of foo without a parameter for d provided by the caller would modify the same default dictionary (equivalent to the second example). 

# this...
def foo(d={}):
    ...

# is equivalent to this
default = {}  # this is a global object!
def foo(d=default):
    ...

You always want to instead use a value like None and check in the function or method body whether a default needs to be supplied, and document the function's behavior accordingly as it cannot be inferred from the type signatures.

# better
def foo(d=None):
    """Does foo

    Args:
        d (dict): some parameter. Defaults to {}.
    """
    if d is None:
        d = {}
    ...

Named tuples

A similar gotcha is using mutable defaults for named tuples.

from typing import NamedTuple

class Foo(NamedTuple):
    d: dict = {}  # no

Again, every instance of Foo that uses the default value will share a reference to a single underlying object. Instead, set the default to None and require consumers to check whether d is None.

Dataclasses

Dataclasses are like named tuples, but better in every way. One feature I recently discovered is the ability to use a default factory to avoid this issue.

from dataclasses import dataclass, field

@dataclass
class Foo:
    d: dict = field(default_factory=dict)

With this paradigm, every time a default needs to be supplied for d, the factory function will be called, creating a new dictionary object each time and side-stepping the gotcha. Additionally, consumers of Foo can always assume that d is a dict and simplify their code accordingly.

python