Python Multinested Closure

After my previous post (that mentions other related posts with similar details) about Python closure introspection (and a bit of internals) I came across a detail that at first seemed strange to me, but that makes much sense (and made me further dive into the implementation).

Let's say we have these nested functions (with 3 levels of nesting). We return the most nested function (inner_2) that traps variables in the most outer function (becoming a closure):

def outer():
    print("outer")
    x = "a"
    y = "b"
    def inner_1():
        # it's using x
        nonlocal x
        x += "b"
        print(f"inner_1: {x}")
        def inner_2():
            # it's using both x and y
            nonlocal x
            x += "c"
            print(f"inner_2, x:{x} y:{y}")
        return inner_2
    return inner_1

in_1 = outer()
in_2 = in_1()

inner_2 is trapping 2 variables defined in outer: x and y. We can see it by checking its __closure__ and the co_freevars in its __code__ object, and the co_cellvars of the outer function code object:

print(f"in_2.__closure__: {in_2.__closure__}.") # 2 cells, for the x and y values>
print(f"in_2.__code__.co_freevars: {in_2.__code__.co_freevars}.") # in_2.__code__.co_freevars: ('x', 'y')

# in_2.__closure__: (cell at 0x78f58889d570: str object at 0x78f58886f930, cell at 0x78f58889d540: str object at 0x5a0cc7144e08).
# in_2.__code__.co_freevars: ('x', 'y').

print(f"outer.__code__.co_cellvars: {outer.__code__.co_cellvars}") # ('x', 'y')
# in_1.__code__.co_freevars: ('x', 'y')

But checking these attributes for the intermediate inner function comes with some surprise:

print(f"in_1.__closure__: {in_1.__closure__}.") # 2 cells, for the x and y values>
print(f"in_1.__code__.co_freevars: {in_1.__code__.co_freevars}.") # in_1.__code__.co_freevars: ('x', 'y').
print(f"in_1.__code__.co_cellvars: {in_1.__code__.co_cellvars}") # () 
print(f"outer.__code__.co_cellvars: {outer.__code__.co_cellvars}") # ('x', 'y')

#in_1.__closure__: (cell at 0x78f58889d570: str object at 0x78f58886f930, cell at 0x78f58889d540: str object at 0x5a0cc7144e08).
#in_1.__code__.co_freevars: ('x', 'y').
#in_1.__code__.co_cellvars: ()
#outer.__code__.co_cellvars: ('x', 'y')

inner_1 is trapping x in its closure, which is normal as it's using it, but it's also trapping y, that it's not using, why? Well, indeed inner_1 is not using y in a direct, visible way, but it needs it, as when the inner_2 function object is created, it needs both x and y for its closure. The cells for x and y are created in the heap when outer is executed. outer creates inner_1 and returns it, so when inner_1 is executed and creates inner_2, outer is long gone, so we need to have the reference to the x and y cells somewhere, to put them in inner_2.__closure__. That "somewhere" is inner_1 closure. So yes, even if inner_1 only works directly with x, it gets y also in its closure.

Discussing this with a GPT you get a nice explanation:

This is sometimes described as “transitive closure capture” or “cell promotion/relaying”: an intermediate function (inner_1) must carry closure cells that it doesn’t itself use, so that functions nested within it can close over them.

In other words: If a nested function needs a variable from an outer scope, every function layer in between must carry that variable as a closure cell, even if those intermediate layers don’t use it directly.

Only the immediate lexical parent can provide the closure cells to a newly created function.

The approach followed by Python for creating its closures is rather different from that of JavaScript, and explains the limitation that I mentioned in this post. In Python the compiler checks if a function closes over variables of its outer scopes, and if so, it sets the co_freevars and co_cellvars of the corresponding code objects and adds the necessary instructions so that at execution time cell objects get created and when the function object is created, its __closure__ can be correctly set, with exactly the cells that it needs. If some "dynamic code" (code compiled dynamically with exec()) tries to access to a variable of an outer scope that had not been trapped by the __closure__ of the function that invokes exec, it can't, as it's not there. In JavaScript this is quite different. eval() has access to any variable of the outer scopes, because indeed all functions in JavaScript have access to all its outer scopes through the scope chain. When a function is created, it gets its [[scope]] property set to the scope (the activation object I think it's called) of its parent function. So if we have a certain level of nesting when defining functions, we end up with a chain of scopes. And the variable look up mechanism will search in this chain if it does not find a variable in the current scope. This is very powerful, but at the same time has serious performance implications. Outer scopes are kept alive regardless of whether the inner functions access to them or not (cause we allow eval to access to them, and we don't know what eval will be evaluating). This also involves extra longer look ups.

Nicely explained by a GPT:

JavaScript keeps the entire lexical scope chain alive, whereas Python collapses scopes into minimal “cell objects” and releases frames as soon as possible.

In JavaScript, every function carries a scope chain because dynamic features like eval() force engines to preserve the full lexical environment at runtime. Python does not need this because its lexical scope is fixed at compile time and not accessible to exec()/eval().

I was wondering how the most powerful and dynamic language that I can think of, ruby, manages this. I have no practical ruby knowledge, so I just asked a GPT, and as expected it follows a very similar approach to JavaScript, keeping sort of a chain of "scopes" that allows eval access to variables in any of them. From a GPT:

Ruby’s closures sit right between Python and JavaScript, but they lean much closer to JavaScript in philosophy:

• They close over entire lexical scopes, not a minimal set of cell-like variables.
• Ruby scopes are runtime objects (not a purely compile‑time fiction like Python’s).
• Blocks, Procs, and lambdas capture the full environment, not a pruned subset.
• Ruby supports eval within a Binding, which preserves the whole lexical + dynamic scope much like JavaScript’s eval.

Python Closure Introspection

I talked time ago about some minor limitation (related to eval) of Python closures when compared to JavaScript ones. That's true, but the thing is that Python closures are particularly powerful in terms of introspection. In this previous post (and some older ones) I already talked about fn.__code__.co_cellvars, fn.__code__.co_freevars and fn.__closure__, as a reminder taken from here

• co_varnames — is a tuple containing the names of the local variables (starting with the argument names).
• co_cellvars — is a tuple containing the names of local variables that are referenced by nested functions.
• co_freevars — is a tuple containing the names of free variables; co_code is a string representing the sequence of bytecode instructions.

And the __closure__ attribute of a function object is a tuple containing the cells for the variables that it has trapped (the free variables).

# closure example (closing over wrapper and counter variables from the enclosing scope)
def create_formatter(wrapper: str) -> Callable[[str], str]:
    counter = 0
    def _format(st: str) -> str:
        nonlocal counter 
        counter += 1
        return f"{wrapper}st{wrapper}"
    return _format

format = create_formatter("|")

print(format("a"))
# |a|

# the closure attribute is a tuple containing the trapped values
print(f"closure: {format.__closure__}")
print(f"freevars: {format.__code__.co_freevars}")
# closure: (cell at 0x731017299ea0: int object at 0x6351ad1bd1b0, cell at 0x731017299de0: str object at 0x6351ad1cd2e8)
# freevars: ('counter', 'wrapper')

A cell is a wrapper object pointing to a value, the trapped variable, it's an additional level of indirection that allows the closure to share the value with the enclosing function and with other closures that could also be trapping that value, so that if any of them changes the value, this is visible for all of them.

def create_formatters(format_st: str) -> Callable[[str], str]:
    """
    creates two formatter closures that share the same 'format' free variable.
    one of them can disable the formatting by setting the format string to an empty string.
    """
    def _prepend(st: str) -> str:
        nonlocal format_st
        if st == "disable":
            format_st = ""  # Example of modifying the closed-over variable
            return
        return f"{format_st}{st}"
    
    def _append(st: str) -> str:
        return f"{st}{format_st}"
    
    return _prepend, _append


prepend, append = create_formatters("!")
print(prepend("Hello"))  
print(append("Hello"))    
# !Hello
# Hello!

prepend("disable")
print(prepend("World"))  # Output: World (since format_st was modified to "")
print(append("World"))   # Output: World
# !Hello
# Hello!

Here you can find a perfect explanation of co_freevars, co_cellvars and closure cells:

Closure cells refer to values needed by the function but are taken from the surrounding scope.

When Python compiles a nested function, it notes any variables that it references but are only defined in a parent function (not globals) in the code objects for both the nested function and the parent scope. These are the co_freevars and co_cellvars attributes on the __code__ objects of these functions, respectively.

Then, when you actually create the nested function (which happens when the parent function is executed), those references are then used to attach a closure to the nested function.

A function closure holds a tuple of cells, one each for each free variable (named in co_freevars); cells are special references to local variables of a parent scope, that follow the values those local variables point to.

If we have a function factory that creates a closure, each time we invoke it we'll get a new function object with its __closure__ attribute pointing to its own object (a tuple), but with __code__ pointing to the same code object. So all those instances of the function have the same bytecodes and metainformation, but each instance has its own state (closure cells/freevars).

The closure "superpowers" that Python features are:

1) As we saw above, ee can easily check if a function is a closure (has cells/freevars) just by checking if its __closure__ attribute is not None (or if its __code__.co_freevars tuple is not empty).

2) We can see "from outside" the values of the closure freevars (the names, the values, and combine both with a simple "show_cell_values" function). And furthermore, we can modify them, just by modifying the contents of the cells in fn.__closure__. It's what we could call "closure introspection".

# combining the names in co_freevars and the values in closure cells to nicely see the trapped values
def show_cell_values(fn) -> dict[str, CellType]:
    return {name: fn.__closure__[i].cell_contents
        for i, name in enumerate(fn.__code__.co_freevars)
    }

def cell_name_to_index_map(fn) -> dict[str, int]:
    return {name: i for i, name in enumerate(fn.__code__.co_freevars)}

def get_freevar(fn, name: str) -> Any:
    name_to_index = cell_name_to_index_map(fn)
    return fn.__closure__[name_to_index[name]].cell_contents

def set_freevar(fn, name: str, value: Any) -> Any:
    name_to_index = cell_name_to_index_map(fn)
    fn.__closure__[name_to_index[name]].cell_contents = value
    
    
def create_formatter(wrapper: str) -> Callable[[str], str]:
    counter = 0
    def _format(st: str) -> str:
        nonlocal counter 
        counter += 1
        return f"{wrapper}st{wrapper}"
    return _format

format = create_formatter("|")

print(f"format cells: {show_cell_values(format)}")
print(f"format 'wrapper' freevar before: {get_freevar(format, 'wrapper')}")
print(format("a"))
# format cells: {'counter': 1, 'wrapper': '|'}
# format 'wrapper' freevar before: |
# |st|

set_freevar(format, 'wrapper', '-')

print(f"format 'wrapper' freevar after: {get_freevar(format, 'wrapper')}")
print(format("a"))
# format 'wrapper' freevar after: -
# -st-

Methods as Closures

Instances of classes and closures feel like 2 competing approaches for certain problems. Instances of classes have state and behavior, but that behaviour is normally splat in multiple execution units (methods). A closure is a single execution unit (a function) that keeps state through the variables it traps (freevars). When a class has a single method, you can model it as a closure (well, a closure factory, so that each closure instance has its own state). Additionally, languages like Python have callable classes, where you have a main/default execution unit (__call__), so they feel closer to a closure :-)

Somehow the other day I realised that in languages like Python or JavaScript, methods of a class can be clousures. How? Well, in Python classes are objects (in JavaScript a class is just syntax sugar for managing functions and prototypes), and we can define classes inside functions, so each time the function runs a new class is created (and returned, so the function becomes a class factory). What happens if a method in one of these internal classes tries to access to a variable defined at the outer function level? Well, it will trap it in its closure. Let's see an example where the format method has access to the pre_prefix variable from the enclosing function.:

def formatter_class_factory(pre_prefix):
    class Formatter:
        def __init__(self, prefix):
            self.prefix = prefix

        def format(self, tx):
            # this method is accessing the pre_prefix variable from the enclosing scope
            return f"{pre_prefix} {self.prefix}: {tx}"

    return Formatter


MyFormatter = formatter_class_factory("Log")
formatter = MyFormatter("INFO")
print(formatter.format("This is a test message."))  

print(f"closure: {MyFormatter.format.__closure__}")  
print(f"freevars: {MyFormatter.format.__code__.co_freevars}") 
print(f"closure[0].cell_contents: {MyFormatter.format.__closure__[0].cell_contents}") 

# Log INFO: This is a test message.
# closure: (,)
# freevars: ('pre_prefix',)
# closure[0].cell_contents: Log

We can write the equivalent JavaScript code and see that it works the same. So JavaScript methods can also get access to variables present in the scope where the class is defined. What this means is that same as regular functions, methods also have a scope-chain (where the freevars will be looked up).

function formatterClassFactory(prePrefix) {
    class Formatter {
      constructor(prefix) {
        this.prefix = prefix;
      }
  
      format(tx) {
        // this method is accessing the prePrefix variable from the enclosing scope
        return `${prePrefix} ${this.prefix}: ${tx}`;
      }
    }

    return Formatter;
  }
  
  const MyFormatter = formatterClassFactory("Log");
  const formatter = new MyFormatter("INFO");
  console.log(formatter.format("This is a test message.")); 
  // Log INFO: This is a test message.
  
  // Notice that JavaScript does not provide direct closure introspection (no equivalent to __closure__), so we can not translate that part from the Python snippet

By the way, it's interesting how for people coming from class based languages the idea that "closures are poor man's class instances" makes sense, while for people coming from functional languages "class instances are poor man's closures". This is discussed here.

Conditional Decorator

After my previous post about decorating decorators I was thinking about some more potential use of this technique, and the idea of applying a decorator conditionally came up. Python supports applying a decorator conditionally using an if-else expression like this:

def log_call(func):
    @wraps(func)
    def wrapper(*args, **kwargs):
        print(f"In function: {func.__name__}")
        return func(*args, **kwargs)
    return wrapper 

@(log_call if debugging else lambda x: x)
def do_something(a, b):
    return a + b
    

That's pretty nice, but at the same time quite limited. We apply or not apply a decorator based on a condition at the time the function being decorated is defined. But what if we want to decide whether the decorator logic applies based on a dynamic value, each time the decorated function is invoked? We can have a (meta)decorator: conditional, that we apply to another decorator when this decorator is applied, not defined. conditional creates a new decorator that traps in its closure the original decorator and a boolean function (condition_fn) that decides whether the decorator has to be applied. This new decorator receives a function and returns a new function that in each invocation checks (based on condition_fn) if the original decorator has to be applied. Less talk, more code:

def conditional(decorator, condition_fn: Callable):
    """
    metadecorator: createa a new decorator that applies the original decorator only if `condition_fn` returns True.
    """
    def conditional_deco(fn: Callable):
        @wraps(fn)
        def wrapper(*args, **kwargs):
            if condition_fn():
                return decorator(fn)(*args, **kwargs)
            else:
                return fn(*args, **kwargs)
        return wrapper
    return conditional_deco

@(conditional(log_call, lambda: debugging))
def do_something2(a, b):
    return a + b    

print(f"- debugging {debugging}")
print(do_something2(7, 3)) 
debugging = False
print(f"- debugging {debugging}")
print(do_something2(7, 3))
print("------------------------")

# - debugging True
# In function: do_something2
# 10
# - debugging False
# 10

Python MetaDecorator

We know that when using decorators in Python you should always use functools.wraps/update_wrapper on the function returned by the decorator. Apart from setting the __name__, __doc__, __module__... attributes of the new/wrapper function with those of the original one, it also adds a __wrapped__ attribute that points to the original function. What it does not do is adding information to the function about the decorator that has been applied. So while we have a way to refer to the original function via __wrapped__, we can not check if a decorator has been applied to the function.

Obviously our decorator could just add a __decorator__ attribute to the wrapper function that it returns, but well, we have to repeat that logic in each of our decorators, and can not do anything with already existing decorators. So the nice way to do this would be having a function (let's call it empower()) that we can apply to an existing decorator, obtaining a new decorator that applies the original decorator and then sets the __decorator__ attribute in the decorated function. This empower function is a decorator factory (receives a decorator and creates a new decorator) and indeed can be applied (at least in some cases, we'll see it later) as a decorator itself when defining the initial decorator, so it could be seen as a sort of meta-decorator (a decorator that decorates and creates decorators).

A function can be decorated by multiple decorators, so shouldn't we better have a __decorators__ attribute with that list of decorators? Well, the __wrapped__ attribute points to the function being decorated in this step (so functools.wraps does not check if the function being decorated has in turn a __wrapped__ attributes). So if we have a chain of decorators we'll have to traverse a chain of __wrapped__ attributes to get to the source function. I mean:

"""
Veryfying that if multiple decorators are applied, functools.__wrapped__ points to the previous decorated function in the chain, not directly to the original function. 
"""
from functools import wraps
def start_call(func):
    @wraps(func)
    def wrapper(*args, **kwargs):
        print(f"Starting function: {func.__name__}")
        return func(*args, **kwargs)
    return wrapper

def end_call(func):
    @wraps(func)
    def wrapper(*args, **kwargs):
        result = func(*args, **kwargs)
        print(f"Ending function: {func.__name__}")
        return result
    return wrapper 


@start_call
@end_call
def do_something(a, b):
    return a + b

# Example usage
if __name__ == "__main__":
    # do_something has 2 levels of decoration
    print(f"Result: {do_something(5, 10)}")

    print("---------------- Unwrapping decorators ----------------")
    unwrapped1 = do_something.__wrapped__
    print(f"Result: {unwrapped1(5, 10)}")
    print("----------------")
    unwrapped2 = do_something.__wrapped__.__wrapped__
    print(f"Result: {unwrapped2(5, 10)}")

    # Starting function: do_something
    # Ending function: do_something
    # Result: 15
    # ---------------- Unwrapping decorators ----------------
    # Ending function: do_something
    # Result: 15
    # ----------------
    # Result: 15

So that's also the approach I've followed here. I add a __decorator__ attribute to each decorated function, and have a get_decorators helper function that will traverse that __decorator__ chain to get all the decorators. My empower decorator provides an additional functionality, if the decorator being decorated does not apply wraps() to the original function, it does it. Let's see an implementation of this empower decorator and the associated get_decorators() function.

def empower(decorator):
    def empowered_decorator(func):
        decorated_fn = decorator(func)
        decorated_fn.__decorator__ = decorator
        # if the original decorator has not used wraps, we add it here
        if not hasattr(decorated_fn, '__wrapped__') or decorated_fn.__wrapped__ != func:
            wraps(func)(decorated_fn)
        return decorated_fn
    return empowered_decorator

def get_decorators(func):
    decorators = []
    while cur_decor := getattr(func, '__decorator__', None):
        decorators.append(func.__decorator__)
        func = func.__wrapped__
    return decorators

Given the previously defined start_call and end_call decorators, we can empower them at the time they are applied to a function, like this:

@(empower(start_call(">>>")))
@(empower(end_call))
def do_something2(a, b):
    return a + b

print(do_something2(7, 3))
print(f"decorators applied to do_something2: {[dec.__name__ for dec in get_decorators(do_something2)]}")
# decorators applied to do_something2: ['intermediate', 'end_call']

Being used at the time a decorator is being applied, rather that at the time when a decorator is defined, the empower decorator works naturally both for decorators that expect parameters and decorators that do not (other than the function being decorated). A decorator that expects parameters does indeed create a new decorator that traps the provided parameters in its closure for being then invoked with the function to be decorated, so in both cases empower ends up receiving a decorator that just expects a function.

We can also apply it when a decorator is being defined, but only for decorators that do not expect parameters (other than the function being decorated itself). For applying it at definition time to decorators that expect parameters, we need a different implementation, that I've called empower_dec_with_params. So all in all we have:

# intended to be used when defining a decorator with parameters
def empower_dec_with_params(decorator):
    def outer_decorator(*args, **kwargs):
        def inner_decorator(func):
            decorated_fn = decorator(*args, **kwargs)(func)
            decorated_fn.__decorator__ = decorator
            # if the original decorator has not used wraps, we add it here
            if not hasattr(decorated_fn, '__wrapped__') or decorated_fn.__wrapped__ != func:
                wraps(func)(decorated_fn)
            return decorated_fn
        return inner_decorator
    return outer_decorator
    
@empower_dec_with_params
def start_call(prepend: str = ""):
    def intermediate(fn):
        @wraps(fn)
        def wrapper(*args, **kwargs):
            print(f"{prepend} Starting function: {fn.__name__}")
            return fn(*args, **kwargs)
        return wrapper
    return intermediate

@empower
def end_call(func):
    @wraps(func)
    def wrapper(*args, **kwargs):
        result = func(*args, **kwargs)
        print(f"Ending function: {func.__name__}")
        return result
    return wrapper 

@start_call(">>>")
@end_call
def do_something(a, b):
    return a + b

print(do_something(5, 10))
print(f"decorators applied to do_something: {[dec.__name__ for dec in get_decorators(do_something)]}")

# >>> Starting function: do_something
# Ending function: do_something
# 15
# decorators applied to do_something: ['start_call', 'end_call']