purify

Pythonic object-mutator transforms as pure functions.

Rationale

This solves a longstanding complaint that I have about Python - there's no Pythonic way to write a pure function that is supposed to return a transformed version of an input instance of a class.

Specifically, Python heavily incentivizes (via its syntax and via utilities such as mypy) doing the following:

def rename_tree(name: str, tree: Tree) -> Tree:
    tree.name = name
    return tree

because the pure alternatives are either expensive:

def pure_rename_tree(name: str, tree: Tree) -> Tree:
    tree = deepcopy(tree)
    tree.name = name
    return tree

or not Pythonic and not type checkable:

def pure_rename_treedict(name: str, tree: dict) -> dict:
    return dict(tree, name=name)

This library is a result of noticing that the first option above is extremely common in our codebases at XOi and enough of a convention to be a best practice. While this usually works as long as we compose these "object transforms" as a linear pipeline where no function ever needs to read the pre-transformed object, it is nevertheless a convention which trades off actual purity for efficiency and readability.

A new decorator, @purify, can be applied to any function where a single one of the arguments will be modified, and by performing a behind-the-scenes shallow copy of that object, allows the object transform to become pure without further ado.

@purify
def rename_tree(name: str, tree: Tree) -> Tree:
    tree.name = tree
    return tree

tree = Tree('Felicity')
dtree = rename_tree('Daniella', tree)

print(tree.name) # 'Felicity'
print(dtree.name) # 'Daniella'

Shallow copy vs deepcopy

In rudimentary tests with objects of size < 400 KB, deepcopy was found to be, not surprisingly, 3 orders or magnitude slower than a shallow copy. Additionally, equality tests on deep-copied and not modified objects are 1-2 orders of magnitude slower than shallow copies, presumably because the shallow copy allows the comparison to do far more id equality checks.

So, deepcopy is expensive. But it's the only way to be sure that your function is actually pure. purify defaults to shallow copy. Why?

Because, as it turns out, it's frequently pretty simple to split mutating functions into 'levels' based on the actual object that they modify, and then decorate each level independently. In most cases, this will mean far fewer Python objects need actual copying, and it also gives you more reusable functions than you would have had if all the levels were present together.

Given:

class Nest:
    num_eggs: int

class Tree:
    # ... (lots of other attributes)
    nests: List[Nest]

This expensive deepcopy approach:

@purify(deep=True)
def lay_in_all_nests(add: int, tree: Tree) -> Tree:
    for nest in tree.nests:
        nest.num_eggs += add
    return tree

Could be replaced with the equally pure, and less copy-expensive:

@purify
def lay_in_nest(lay: int, nest: Nest) -> Nest:
    nest.num_eggs += add
    return nest

@purify
def lay_in_all_nests(lay: int, tree: Tree) -> Tree:
    tree.nests = [lay_in_nest(lay, nest) for nest in tree.nests]
    return tree

How to visually parse shallowly pure functions

Some effort is required to use shallow-copy functions properly, whereas deepcopy makes your function trivially pure. How to focus that effort?

A good rule of thumb is that the object being purified must only ever be referenced with a single dot (.), e.g. tree.nests, and usage of that dotted name must either be read-only or direct assignment to that name. E.g., tree.nests[i] = foo is a no-no, because the left-hand-side of the statement is not the bare name nests, but something that directs its activity into the list itself.

Advanced Usage:

Argument name

It's highly recommended to follow a convention where the object that you're mutating is the last positional argument to your function. This is generally better for the composition of many partially-applied functions transforming the same object.

That said, if you have a desire to specify which argument is to be shallow-copied, you may do so by calling the decorator with the first positional argument being the name of the function argument you want purified.

Deepcopy

As above, if you have a need for deepcopying, you need only to pass deep=True to the decorator.