Make a python function definition stick!
A terrible hack that ends up being surprisingly useful
Posted on November 27, 2022

Sometimes we have a function that takes a while, and we want to use just-in-time compilation to make the function faster. Numba will compile the function lazily in this case: only when it is first called will the function be optimized.

from numba import jit

def function_that_takes forever(*args, **kwargs):
    # do some long computation
    return result

This is an excellent way to reduce compute time when making a repeated calls to slow function. Excited with this technique, we implement some long running function in some type of Python REPL (my favorite at the moment is ptpython) we are analyzing data in and slap the jit decorator on top.

def long_running_repl_function(*args, **kwargs):
    # Now our notebook will no longer take forever to run!
    return result

We use this new function with reckless abandon in our code, and it significantly reduce the time it takes our code to run in our REPL. At some point we press the up arrow to get to our old blocks of code, rerun then with enter, and…

and it now takes forever to run the function the first time. What happened?!

Functions are redefined every time

When using a REPL, the code in a block is rerun every time we execute that block. The definitions in the block being run are assigned to the global namespace in the running python interpreter. Put another way, we can think of running a block as the following pseudocode.

# Before running a block, we can find all globally defined
# variables in the `globals()` dictionary. This will include
# a pointer to all of our defined functions, plus potentially a
# bunch of stuff that our REPL may keep track of like `In` and
# `Out` blocks in something like ipython or ptpython.
before = globals()

# Take the source code in the block as a string called `block_str`.
# Anything defined in the block string will be merged into the 
# `globals()` dictionary.

# The globals dictionary should now be modified with whatever was
# defined within `block_str`.
after = globals()

# The following equation relates the two dictionaries
# (minus some jupyter bookkeeping)
#     after = before | top_level_definitions(block_str)
# where `|` is the dictionary merge operator from Python 3.9 and
# `top_level_defintions` returns the top level definitions in the
# string along with the associated function pointers created by `exec`.

The reason the jit decorator does not seem to work is that when we run a block, a new function with the same name gets created and replaces the existing function in the global namespace. We lose track of our first jitted function!

But all is not lost!

For the jit decorator to working even after running the block that defines the desired function multiple times, we need to keep track of the original function with the jit attached and ignore the new function created when the block is run again. Conveniently, the globals() dictionary gives us access to the global namespace, enabling us to see if a function with a given name already exists and hold onto the original definition instead of the new (likely equivalent) definition. The name of a function can be looked up through the __name__ property, so we can define the following decorator1 to find the existing function and keep it.

def sticky_definition_first_try(f):
    # Here we try to find the function in the global
    # namespace. If it does not exist, return the 
    # newly defined function.
    return globals().get(f.__name__, f)

Now we can decorate our function with sticky_definition_first_try once it is jitted to prevent subsequent reruns of a block from clearing out our optimized function.

def long_running_repl_function(*args, **kwargs):
    # Now our notebook will no longer take forever to run!
    return result

Now we can go back to merrily coding in our REPL without worrying about recomputing expensive results.

At some point, we realize our long_running_repl_function has a bug. No problem, we can go back and fix the function and go back to the task at hand.

def long_running_repl_function(*args, **kwargs):
    # Now our notebook will no longer take forever to run!
    # We were missing a constant 5 from our results. Fixed now!
    return result + 5

We rerun the block defining long_running_repl_function and all the blocks that need this function. But we are getting the same resulting buggy results as before. Hmm, what is going on?

Since we only define a function once, we can’t modify it!

Our strategy for preventing a function from being redefined is a tad too strong: it only keeps track of the first definition of the code, meaning we can never alter that definition. It would be amazing if we could write perfect code so that we would not need to redefine a function, but alas we are human. What the sticky_definition function should really be doing is detecting if the function meaningfully changes, and if it does, to accept the new definition (invalidating the old jit in the process).

What we would really like is a representation of only the computational bits of the function, which we could then use to compare the existing implementation of a function from the redefinition from a function. Luckily we can extract the abstract syntax tree (AST) for a function in python using the ast module in the standard library. The AST is the data structure that represents the program after it has been parsed from its text origins, retaining only the computationally relevant bits.

We can define our sticky_definition function to compare the AST of a functions and its redefinition and only accept the redefinition if it does something meaningfully different2 (with generous help from StackOverflow).

import ast
import inspect
import hashlib

# From
# with light editing
def _remove_docstring(node):
    if not (isinstance(node, ast.FunctionDef) or
            isinstance(node, ast.ClassDef)):

    if len(node.body) != 0:
        docstr = node.body[0]
        if (isinstance(docstr, ast.Expr) and
            isinstance(docstr.value, ast.Str):

# From
# with light editing
def hash_function(func):
    func_str = inspect.getsource(func)
    module = ast.parse(func_str)

    assert (len(module.body) == 1 and
            isinstance(module.body[0], ast.FunctionDef))
    # Clear all the doc strings
    for node in ast.walk(module):

    # Convert the ast to a string for hashing
    ast_str = ast.dump(module, annotate_fields=False)

    # Produce the hash
    fhash = hashlib.sha256(ast_str)
    result = fhash.hexdigest()
    return result

def sticky_definition(f):
    # function was not defined before, just return the new function.
    if f.__name__ not in globals():
        return f

    global_f = globals()[f.__name__]

    if hash_function(f) == hash_function(global_f)
        # Return the original function, ignoring the new definition
        # of `f` ONLY IF the source code is unchanged.
        return global_f
        # Otherwise, the function does not already exist with the same
        # source code and we should assign the function to the
        # global namespace.
        return f 

And with this version of sticky_definition we can

  • use the advantages of a stateful decorator like jit in our Python REPL, even after rerunning a block,
  • reduce the number of cases where we have no jitted result, and
  • are immune to syntactic changes like docstrings or comments causing a jit miss.

Is there a more general way to detect a “meaningful” change?

A natural question after using the AST to compare two functions is whether this method also redefines a function too often, much like our source code comparison method was too strict with its understanding of what a “meaningful” change was. Is there a more general definition of a “meaningful” change that we should be looking for?

One more general definition is if our original function and its redefinition always return the same output for a given input, then we can consider no meaningful change has been made in the redefinition. For example, we can define a function in two ways that produce the same result.

def definition_one(a: int) -> int:
    return a + 5

def definition_two(a: int) -> int:
    return 5 + a

Since these two definitions should generate the same jit, we can consider them equivalent. This definition of equality is called function extensionality: if for all inputs x we can show that, for two functions f and g that f(x) = g(x), then we can say that f and g are extensionally equal. Our AST comparison we used for function equivalence was an intensional equality: two functions were equal if they were syntactically equal for a given definition of syntaxes being equal. In this case, we defined the syntaxes as being equal if the AST was the same.

Function extensionality seems to be a good way to decide if two functions are equivalent, so why did we not use that? Well there are two main reasons I can think of. The simpler problem is that if we redefine the function to be far more computationally efficient, then we would be ignoring this more efficient definition since the input/output mapping is the same.

More importantly though, in general it is difficult to determine if two functions are extensionally equivalent except for in very specific cases, especially in languages such as Python3. One could attempt to recognize some ASTs as equivalent through rules that recognize some equivalent transformations. For example, a rule could detect that a + b is the same as b + a. It would be difficult to come up with a set of rules that encapsulated the right definition of a meaningful difference between two ASTs, likely taking far more time than just rerunning the expensive computation. In addition, I would argue that when we edit code that changes the AST that we are usually making meaningful changes, so we are unlikely to redfine a function more than we really want to in practice.

Should one really want to extend the definition of equality, then Satisfiability modulo theories (SMT) solvers can be used to symbolically reason about programs. Using such tools, one can determine equality of two functions by asking if there is some symbolic assignment applied to both functions that returns an unequal result. Tools such as Crux from Galois or the work from the Languages, Systems, and Data Lab @ UCSC use this technique to prove that a function implements a certain specification, such as producing equivalent outputs as a less efficient function. Normal testing is insufficient in this case as it is often not possible to try every input to a function, meaning it is possible to miss a bug arising from an edge case.4

Should we do this?

This approach to preserving a jit is definitely a hack: we are altering how functions are assigned to a global namespace. Anytime code that pulls out the ast module, the inspect module, or the globals function is likely doing something that is probably gross—if we are using all three, then we should definitely rethink what we are doing! This idea should not be used in actual production code.

However, for prototyping (including in places like a REPL) this method enables us to quickly enable decorators with state (such as jit) without trying a more canonical solution such as persistent storage. Sometimes coming up with that robust saving state challenge can take quite a while in its own right!


This type of idea has been done a few different times in the public literature.

Comments, questions?

If you have any comments or questions, feel free to do one of the following.


I would like to thank Patrick Redmond and Spencer Poff for reviewing drafts of this article.

Want to run the code in this blog post?

A python environment can be spun up using nix shell in this posts root directory. Play around with different definitions for comparing functions and see what kinds of tradeoffs you encounter.

  1. A decorator is just a higher order function (a function that takes another function as an argument).↩︎

  2. See references for examples of performing this type of AST comparison.↩︎

  3. There are some programming languages such as Dhall which have a more generic way of comparing functions through strong normalization.↩︎

  4. Thanks Patrick Redmond for suggesting expanding the section on ways to evaluate function equivalence.↩︎