Python's philosophy of what closures close over

October 7, 2011

A commentator on my last entry on closures said:

I don't not fully understand why closures could not capture bindings upon being defined, though: [...]

As your first example shows, this is what many people intuitively expect, it seems. So, do you know why this behavior was not chosen?

I think the real answer is 'because how Python does it is how languages usually do it'. Let me elaborate on that.

When you add closures to a language, you are faced with a design choice (or a philosophical question). Consider the following code:

def dog(a):
  def cat(b):
    return a + b
  return cat

Here a is a 'free variable' in cat, and when the closure is created it must be bound to something. The question is what it binds to: does it bind to the variable a, or does it bind to the value of a at some point in time?

(Given a suitable choice of when a closure is created, I don't think that there's a right or a wrong answer here, just different ones.)

Like most languages with closures, Python has picked the first choice; free variables are bound to variables, not to values. There are pragmatic reasons for choosing this, including that it is most compatible with how global variables are treated in ordinary functions, but ultimately it is a choice. I don't think there's anything in Python's other semantics that require it.

Making this choice creates a second one, which is what a variable's 'lifetime' is; when does a stop being a and become another variable (even one with the same name)? This is not quite the same as the scope of a, which is conventionally more or less when the name a is defined. To see the difference, consider the following pseudo-code:

nums := [];
for (var j, i = 0; i < 100; i++) {
  j = derivefrom(i)
  nums.append(lambda x: j+x);

In this code the scope of j is the for loop and the name is undefined outside of it. However we have two choices for the lifetime of j. We could say that j is the same variable for every pass through the for loop body or we could say that each separate pass through the loop body creates a logically separate version of j (which we could call j0, j1, and so on). In the latter case each lambda binds to a separate version of j, the version that was live during its loop. In the former case all lambdas bind to a single version (and the version has whatever value came back from 'derivefrom(99)').

This example probably seems artificial. Now consider the same question recast as:

nums = (lambda x: x+i for i in range(0,100))

The scope of i is the inside of the generator expression and it is undefined outside of it, but does every separate invocation of the generator's body use the same i or does each invocation get a logically separate i? Again this is a design choice and I think that you could answer either way. Python chooses to say 'every invocation of the generator's body uses the same i'. What happens to the lambdas then follows logically.

(If nothing else, this makes the interpreter's implementation simpler.)

PS: with very careful use of scopes you can make what I'm calling variable lifetime here into scope lifetime, even with each loop body getting its own version of j. However, it requires a somewhat convoluted definition of how loops and loop index variables work. (And that is well beyond the scope of this entry.)

Sidebar: a pragmatic problem with 'bind to value' in Python

The problem is that def is an executable statement. This means that the closure is actually created in dog not when the line 'return cat' is executed but when 'def cat ...' occurs in the source. You can see this directly by disassembling the bytecode in a version of dog where there's a statement between the definition of cat and the return.

This means that if closures bound to values instead of variables you might well have to move inner function definitions down in your function so that they were intermixed with actual code, because you would need to insure that they were only defined (and the closure created) after all of the variables they needed had been set up. And (as I saw mentioned somewhere) you couldn't have mutually recursive closures without awkward hacks.

Comments on this page:

From at 2011-10-07 04:08:30:

Thanks for clarifying the concepts and the choices at hand.

Your sidebar is what sold me to the "bind to variable/name" philosophy. It is in fact convenient to define functions that use names that are not yet defined, and I like the fact that closures are no different.

-- EOL, commentator on your last entry on closures :)

From at 2011-10-07 06:17:01:

I can't think of any language that makes the design choice to bind inner references to an outer variable to the value of that variable at the point of definition and not to the variable itself. The closest I can think of is that java sidesteps this problem by allowing non-static inner classes (closest thing java has to closures) to only refer to variables in a surrounding scope that are declared final. (Meaning that the variable isn't really a variable in a sense, since it can't vary after declaration) I suppose you could say haskell follows a similar side-stepping of the issue by effectively having every variable final.

Written on 07 October 2011.
« Thinking about event loops versus threads
My new Linux machine for fall 2011 (planned) »

Page tools: View Source, View Normal, Add Comment.
Login: Password:
Atom Syndication: Recent Comments.

Last modified: Fri Oct 7 02:42:01 2011
This dinky wiki is brought to you by the Insane Hackers Guild, Python sub-branch.