I've got a little project where I want to build a document from a template, and some data that are stored in a JSON (or YAML) file. This is pretty straighforward, but I've got an additional requirement: I want to make documents in multiple languages, and I want to specify the various translations of a string in the JSON file itself, like so:

{ "en": "This is a sentence in english",
  "fr": "Ceci est une phrase en français"}

This seems quite easier, so I started coding straight away, and I arrived to this result:

def localize(lang, data):
    if isinstance(data, list):  # If this is a list, recurse into the list
        return [localize(lang, x) for x in data]
    if isinstance(data, dict):  # If this is a dict, do what we must
        if lang in data:
            return data[lang]
        else:                   # Otherwise, do nothing
            return data
    return data

Can you see the problem with this? I failed to recurse in the case of the dict. Of course, this is easily fixable: just replace the first return data with return {k: localize(lang, v) for (k, v) in data.items()}. But this is a common mistake, and it stems from the fact that we mixed our traversal code, and the actual transformation we wanted to make. Would it be possible to write our code in a way where no mistakes can be made?

Complicated words to the rescue!

I've been spending quite some time with haskell lately, and I recently learned about recursion schemes: a generic way to handle recursive data structure. In python, we have a generic way to handle iterative data structure: iterators. They let you handle a list, a tree, or anything with elements the same way: with a for loop. Recursion schemes are the same idea, but for recursive data.

A recursive data type is a type where one value might contain another value of the same type. For example, a JSON value is recursive, because it can contain another JSON value (in a list or as the attribute of an object).

In haskell, if a type can contain something (whatever it is), you call it a functor, and you give it a function called fmap, which execute a given function on all contained items. Let's implement it for our JSON values!

def fmap(f, data):
    if isinstance(data, list):
        return [f(x) for x in data]
    if isinstance(data, dict):
        return {k: f(v) for (k, v) in data.items()}
    return data

Now that JSON values have a fmap function, we can use it to implement our generic recursive traversal. We want to traverse our data to build a new value along the way: this particular kind of recursion is called a catamorphism. Fear not, it's a rather simple implementation for a complicated word.

def cata(f, data):
    # First, we recurse inside all the values in data
    cata_on_f = lambda x: cata(f, x)
    recursed = fmap(cata_on_f, data)

    # Then, we apply f to the result
    return f(recursed)

When you pass a function f to cata, it is called at each level of the recursion, and is passed a copied version of the data where each contained value x is replaced with cata(f, x). It means that f doesn't have to worry about the recursion, when f is called, the recursion has already been done.

For example, say we want the sum of all the integers in a tree (which we model as a nested list of lists). We need to write a function that only sums one level, since the recursion is done by cata. The function receives either a value that is not a list, and just returns it, or a list were values have already been summed.

def sum_one_level(data):
    if isinstance(data, list):
        return sum(data)
    return data

Now, we can call cata(sum_one_level, [[[1, 2], [3]], [4]]), and the result is 10.

Back to the localization

So, we now have a cata function that does the recursion for us, and does it well. To recreate our localization function, we need to define a function that only localizes one level. It takes a JSON value, and localize it (but without any recursion, since the recursion is handled by cata).

def localize_one_level(lang, data):
    if isinstance(data, dict) and lang in data:
        return data[lang]
    return data

We can then define our new localize using this function.

def localize2(lang, data):
    localize_one_level_on_lang = lambda x: localize_one_level(lang, x)
    return cata(localize_one_level_on_lang, data)

Here we go! I've defined a function to recurse into JSON data in a generic way, and I'll never have to write a recursive function for JSON data ever again. I'll just need to define functions that work on one level, and that's it. And I can do that for any other recursive data type: I'll just need to create another fmap implementation for them.