Veit's Blog

A Bad Idea


TL;DR because the introduction is a bit long: We’re going to play around with Python’s metaprogramming facilities to build a plug-in system for Python’s parser!

In the summer of 2016, I flew to Bilbao to attend EuroPython. I was still working on zepto, and used some of my downtime to mull over some of the concepts that I wanted to put in the language, or had already implemented. I told a lot of people about it, too, and many seemed very interested, if only in all the ideas that I had.

Naturally, I eventually started to think about how I could get some of the concepts down in Python in the hopes that people would then have an easier time understanding what they meant. Most of the features seemed impossible to graft onto the language without touching the interpreter itself, which would make it both tedious for me to implement and for the curious to build.

Then I thought of one of my favorite talks at the time, in which David Beazley talks about metaprogramming in Python for three hours. Among other things, he builds a loader and importer for XML files that follow a certain schema, like a declarative way to define classes. While this is certainly at least a little silly, in my mind I connected that to a feature that zepto “borrowed” from Racket: the #lang reader shorthand.

This concept blew my mind when I first encountered it: by writing #lang <name> at the top of any Racket file, you tell Racket to first pass the contents of that file to a transformer that says it can read in that source code and transform it into regular Racket1.

This might all sound kind of weird and technical, but what this is saying is that you can plug in your own parser before Racket’s!

That was super cool to me, and I added a simplified version in zepto. Because the procedure that loads source files is written in zepto itself, I was able to just add a quick check to the loader that sees whether the source file starts with a #lang directive and if any function is registered that promises to be able to transform that code. It is then passed onto the function, which is expected to return regular S-expressions—another lisp-y construct—that we can then expand and evaluate as usual2.

So I decided to whip up a version of that for Python and demo it in a lightning talk. Sadly, all of the spots for talks were already taken, so I shelved the project and moved on.

A few days ago I realized that this project might actually make an interesting instructional blog post, however, and here we are. What we’re going to do is write a few Python classes that will enable us to have a crude version of the #lang directive for Python files!

The source code for this project is on Github, if you want to follow along.

Just to be clear, I don’t advertise you use this for anything other than your personal enjoyment. While this concept works well for a language like Racket, it doesn’t work super well in Python.


We’ll begin by defining an API. The module will be called a_bad_idea, and anyone who wants to implement a new language will first have to register their parser function. Let’s see how we would do that for a JSON file, for instance:

import json
import a_bad_idea

def json_parser(code, module):
  module.value = json.loads(code)

a_bad_idea.add_implementation("json", json_parser)
Fig. 1: A simple JSON loader.

Note that the module we get back in this case would have one definition in it named value, which is a little weird, but could potentially be convenient. Imagine having a configuration file in JSON that you want to read in; just importing it would be so much simpler, right?

We will also check that the file extension matches the name in the directive. This might seem redundant, but while we’re learning about the black magic of importers, why not do some extra credit work, right?

The code

Alright, let’s get to writing code! First of all, let’s think about how we could make that API possible. I’ll go for the simplest solution: a global dictionary in our namespace:

LANGS = {}

def add_implementation(name, parser):
  LANGS[name] = parser
Fig. 2: The code for adding languages.

That was easy! Now we can register parsers, which is great, but they won’t actually do anything yet. For that, we’ll need to talk about how finding and loading modules works in Python. The sys module has a property called meta_path, which are essentially module finders. They need to have a class method called either find_spec (since Python 3.4) or find_module (in prior versions), which will try to find the module that is being imported and then return another class that actually imports it. But we’re getting ahead of ourselves here! Let’s first insert our finder into that list:

import sys

class Finder:
  def find_module(cls, fullname, path):

sys.meta_path.insert(0, Finder())
Fig. 3: A skeleton for our finder.

Alright, so as it currently stands our finder will not do anything. We’re implementing find_module because it’s called as a fallback in modern versions of Python as well, which means that it’s the only function that works across versions. If you want to make this a little prettier, you could also implement find_spec and just pass the arguments on to find_module.

Now we need to make our finder actually do things. It will need to go through all the directories in the path, find the first module that is eligible, and return it. If we don’t find anything, Python will resort to the other loaders in the list.

ìmport os

class Finder:
  def find_module(cls, fullname, path):
    # we need to get through the path
    for dirname in sys.path:
      # get the filename without extension
      basename = os.path.join(dirname, fullname)

      # iterate through our implementations
      for extension in LANG:
        # build the real filename
        fname = '{}.{}'.format(basename, extension)

        # and make sure it’s a file (and not a directory)
        if os.path.exists(fname) and not os.path isdir(fname):
          # check if we have the right directive and
          # return our loot
          lang = _get_lang(fname)
          if lang and lang == extension:
            return Loader(fname, lang)
Fig. 4: A real finder.

That’s a bit of a mouthful, but it’s mostly comments, and while the code uses some lesser known functions of the Python standard library, it should be somewhat straightforward. sys.path is a list of the directories that Python looks into when you import a module.

There are two things in the code in Figure 4 that we need to implement before we’re done. Firstly, we need to check for the directive and then we need to implement a loader. Because the latter is a little more work, let’s start with the former.

def _get_lang(filename):
  # we get the first line
  with open(filename) as f:
    first_line = f.readline()
  # check it for the directive
  if first_line.startswith('#lang '):
    # and get the rest of the line, stripping off the newline
    words = first_line.split(' ')
    return ' '.join(words[1:])[:-1]
Fig. 5: Checking for the directive.

As we suggested earlier, this step is basically redundant, and just there for show. We’re also not parsing anything yet, operating at a level of files and strings, which might come as a surprise.

Now let’s fiddle with some modules, shall we?

import imp

class LangLoader:
  def __init__(self, fname, lang):
    self.fname = fname
    self.lang = lang

  def load_module(self, name):
    # first we check whether we already have that module
    if name in sys.modules:
      # if so, just reuse that
      mod = sys.modules[name]
      # if not, we’ll create one and add it to the cache
      mod = imp.new_module(name)
      sys.modules[name] = mod

    # set some minimal metadata
    mod.__file__ = self.filename
    mod.__loader__ = self

    # get the parser
    parser = LANGS[self.lang]
    # execute and return the module
    parser(_get_code(self.fname), mod)
    return mod
Fig. 6: A custom loader.

Again, this might seem like a lot, but it’s actually quite straightforward. We use sys.modules to store and retrieve imported modules. It’s generally a bad idea to just fiddle with the internals of Python willy-nilly, but in the module finder and loader there’s just no way around it, and it’s actually a little bit fun as well.

Getting and executing the parser is a little anti-climactic; we just pass in the freshly created module and the code and hope for the best. But it makes sense if you think about it: there’s actually not that much mechanic in a loader if it doesn’t have to parse anything. It’s just setting the playing field for our user-supplied parsers and that’s it.

We’re not quite done yet, however. For completeness’ sake, we’ll also have to implement _get_code, which is a very simple function that just reads the file and strips of the first line, because we don’t want to have the #lang directive in the string we pass to the user.

def _get_code(filename):
  with open(filename) as f:
    contents =
  return "\n".join(contents.split("\n")[1:])
Fig. 7: Preparing our source code.

And that’s it! In just over 50 lines of code, we added a plugin-based parser system to Python! Now all that’s left to do is write some parsers, but I’ll leave that as an exercise to my beloved readers.


While we weren’t quite able to replicate the power and wonder that you feel when writing languages using Racket, we were able to build an extension to Python without ever touching the core of the language, and I think that’s quite beautiful.

If you want to learn more about writing languages using Racket, I recommend checking out Matthew Butterick’s fantastic Beautiful Racket, in which he guides you through the process of writing a few little languages in Racket. It’s quite fun, and I had a blast reimplementing the languages from the book in zepto back in the day!


1. This is quite a bit of a simplification, but I didn’t think the actual mechanics actually matter that much. If you want to dive a bit more deeply into how it all comes together technically, I suggest looking at the Racket documentation!

2. I want to note here that this is much, much simpler than the machinery in Racket, which relies on reader macros. Racket also has tremendously more tooling around the concept, which shouldn’t come as a surprise.