Python Worlds

Last year, I was introduced to a paper via Lambda the Ultimate about worlds, a language construct which allows one to control the scope of side effects while programming.

Worlds allow you to capture the current scope of a program in a first-class way. All updates to the current state (i.e. local variables, global variables) happen in a non-commiting way. In other words, you can back out of any changes at any time.

Consider the following example (taken from the Warth paper):

A = thisWorld; // thisWorld is always the current world
r = new Rectangle(4, 6);

B = A.sprout(); // sprout creates a new world with it's parent set to A 
in B { r.h = 3; } // side effects within this `in' occur only in the world B.

C = A.sprout();

in C { r.h = 7 }; // in A's world r.h = 6 still.

C.commit(); // only now does r.h = 7 in world A.

If you follow along in the comments I’ve appended to the example, you’ll start to see why this idea is interesting, even from this little example.

The astute Scheme programmer, however, will notice almost certainly that this construct could be created with call/cc, which is certainly true. The problem with this fact is that not all programming languages are Scheme (unfortunately), and of course not all languages support first-class continuations.

The question I asked myself, however, is this: Can I hack worlds into Python? To which I came up with the short answer after some thinking, sort of.

I guess I should explain what’s going on more clearly in the example above. The first thing to note is that A represents the current scope; the current state of all variables in the program. Sprouting a new world from an existing world means that any changes that occur when using the sprouted world, do not affect the world who sprouted the current world, unless the new world commits the changes made to the original world.

Which is to say, changes that occur in an in block acting on world X do not propagate to the parent (the world X was sprouted from) of X unless X.commit() is called.

Enter Context Managers

A few months ago, I wrote a blog post about Dispatching With “with”, in which I explained context managers in Python, and how they can be exploited to create a less separated mapping from URLs to request handlers (something that definitely has its place in the small web-app world).

The basic idea of this was that in the __exit__ method of the web- application object, the current frame was inspected and references to functions that represent HTTP methods would be collected, stored and tied to the last regular expression passed to the expose method in the application object. This simple solution allowed us to express a web application succinctly like so:

app = web.application()
with app.expose('/'):
   def get():
       return "Hello World"
app.run()

For worlds, I also exploit context managers, though mostly for the in-like syntax, and for managing the current thisWorld variable.

The quick1 solution that I came up with for implementing worlds can be used like so:

with Universe(): # establishes new world, assigns to local variable `thisWorld'
   thisWorld.r = True # must assign _directly_ in the world. LIMITATION
   new = thisWorld.sprout()

   with new:
       new.r = False

   with new.sprout():
       thisWorld.r = 15
       thisWorld.commit() # now new.r is 15, but the original r is still True

   print thisWorld.r # => True

   new.commit()

   print thisWorld.r # => 15

   thisWorld.commit() # have to commit to the actual scope LIMITATION

   # r is now part of the local variables where this universe exists

   print r # => 15

Looking at this example, it’s already apparent that the Warth implementation of worlds is superior, just in the amount of code needed to take advantage of it. You might also see that I didn’t even attempt to port the rectangle example from above. That is because there isn’t anything smart going on under the hood when it comes to container objects (such as lists, tuples, objects, dicts), and I’m not yet sure how to get there.

With simple immutable objects such as booleans, integers and strings, using copy-on-write semantics works wonderfully. Then, on commit of the world, the code just copies all of the changes into its parent. I haven’t tackled the case of mutable container objects just yet, as there are complications in the API2, as well as the implementation.

The interaction with this is sort of annoying though. In order to take advantage of worlds in Python, you have to touch virtually every line of code in the function you are trying to worldize, because you must assign explicitly to a world. The world’s context manager sets up thisWorld for you, but you still have to do thisWorld._variable_ to get any sort of benefit.

My inclination is to get into some bytecode hacking to modify all assignments within the with block to be assignments to thisWorld automatically, but bytecode hacks are neither pleasant to maintain, nor are they portable across implementations.

It’s also possible in the Warth version to worldize functions and any other first class object. Maybe the solution is simple and I just haven’t seen it yet. Whatever hacks, that I come up with though, will be just that, hacks, as there is no easy way to add worlds to Python in the same way that Warth added them to JavaScript3.

We are in an age of programming where mainstream programming languages are unable to adapt to our needs as programmers. We are unable to bend them at our will like we can with Scheme, Lisp and even Clojure. Attempts to bring about change on this front have not been met with enthusiasm from most groups. Whether it’s a lack of marketing, evangelism or just that the general population doesn’t view unbendability as a problem, I’m not sure. But, I for one like the idea of being able to easily add worlds, and other ideas, as true language features to languages that by practicality, I’m forced into using. That would make me a much happier, and effective programmer.


  1. By quick, I do mean quick. This was 2 hours of work and sketching. Surely there is lots of work to be done to make it a true solution.

  2. The same strategy could be used as for simple values like booleans, if the API used a method, say assign instead of the more natural assignment operator. Consider, thisWorld.assign('obj.height.inches', 30) vs. thisWorld.obj.height.inches = 30.

  3. The Worlds prototype was written in [OMeta][13], which is a solution to the “unbendable” languages problem. Note: I didn’t attempt to write worlds in PyMeta, but, it may be possible to do.

— 2009-10-01