Hacker News new | past | comments | ask | show | jobs | submit login
Python Anti-Pattern (valinsky.me)
37 points by valinsky 7 days ago | hide | past | favorite | 25 comments

A bit of a nit on the step-by-step rundown of what happens. At step 4 it says the Lambda “dies” but I think this is misleading. The response has completed, but the Lambda has not died. In fact the process is still running which is why the mutable state issue presents.

Even if the process died, the container stopped, and the Lambda went “cold”, I wouldn’t even call that “dead”. It’s just cold. Dead would mean totally failed or decommissioned, I think.

Agree. Updated that part. Thanks for pointing it out.

About half the bugs I fix is due to unexpected interactions with mutable state.

Now, I am usually the creator of the bugs, and it might be that I completely lacks any intellectual rigour with regards to mutable state, but it is re-assuring that I am bot the only one.

This specific behaviour of python has always struck me as completely idiotic.

> This specific behaviour of python has always struck me as completely idiotic

I'm curious, what would be your proposed solution? Make a special case exception for when the expression that constitutes a function default is evaluated? What would that exception look like?

My solution would be to fix it and break backwards compatibility: Any default value is evaluated anew on each function call without the specified value. Then you CAN mutate it to your hearts desire without impacting subsequent invocations. If you want to keep hidden mutable state between function invocations you should use a closure.

    def banana(parts = []):
Calling that procedure without parts would mean parts would be bound to a NEW empty list, not the same old.

Every new pyhton programmer is bitten by this sooner or later. Most people seem to think it is stupid, and relying on it seems, to me, to go against the python mantra of "being explicit", where explicit would mean either using an instance variable in the case of OO or a closure.

Indeed, when I asked for potential solutions, I thought it would be ... weird, to evaluate default args at call time, but it turns out that many other languages are like that[0], and the wierdness was my subjective perception shaped by my Python-heavy background.

Regarding breaking backward compatibility in a future release, I'm not qualified to weigh in on the design decision of balancing between the disadvantages of breaking backward compatibility and the advantage of more intuitive semantics.

[0] https://github.com/microsoft/pyright/discussions/2306#discus...

One useful case for the existing is to eagerly bind variables:

Here's a stupid example. Notice in the first case printing i doesn't grab the value of i when the lambda is defined. The second case is likely what you want.

  >>> lambdas = [lambda : print(i) for i in range(2)]
  >>> for l in lambdas:
  ...   l()
  >>> lambdas = [lambda i=i: print(i) for i in range(2)]
  >>> for l in lambdas:
  ...   l()

Oh my, that's awful. Even in python there are better ways to close over a value :)

I haven't written python in many years, but in the following example it is at least clear what is going on:

   def make_printer(n):
       def printer():
       return printer
   [make_printer(i) for i in range(2)]

Possibly something like this? (Code for illustration purposes only, of course -- please no one use this.)

    class AntiAntiPattern:
      2     """A quick and dirty attempt to mock a Python class instance that 'undoes' the antipattern 
      3        described here: 
      4          https://docs.quantifiedcode.com/python-anti-patterns/correctness/mutable_default_value_as_argument.html)"""
      5     __mutable_args__ = {} # TODO: a real implementation would probably use a WeakMap of some sort
      6     def __init__(self, a, b=None, c={}, d=set()):
      7         self.a = a
      8         self.b = "foo" if b is None else b
      9         self.c = c
     10         self.d = "bar" if d is None else d
     11     def __getattr__(self, k):
     12         return self.__mutable_args__[k] if k in self.__mutable_args__ else super().getattr(k)
     13     def __setattr__(self, k, v):
     14         if v in self.__init__.__defaults__:
     15             self.__mutable_args__[k] = v
     16         else:
     17             self.__dict__[k] = v
     18     def __repr__(self):
     19         return json.dumps(dict(self.__dict__, **self.__mutable_args__), default=str)
    >>> t = AntiAntiPattern(4, d = 'hello!')
    >>> t.__mutable_args__
    {'b': None, 'c': {}}
    >>> t.__dict__
    {'a': 4, 'd': 'hello!'}
    >>> t
    {"a": 4, "d": "hello!", "b": null, "c": {}}

Edit: Oops, fixed a tiny logic bug in __init__. Trying to ween myself off of the 'a = a or "my_a_default"' syntax as instructed by previous commenters in this thread. :)

Edit2: It dawned on me that this doesn't actually resolve the 'gotcha' unless `v` is replaced with `deepcopy(v)` on line 15; otherwise the same exact problem rears its head:

    >>> t = AntiAntiPattern(3, {})
    >>> t.d
    >>> t.d.update([3,4])
    >>> t
    {"a": 3, "b": {}, "c": {}, "d": "{3, 4}"}
    >>> t2 = AntiAntiPattern(4)
    >>> t2
    {"a": 4, "b": {}, "c": {}, "d": "{3, 4}"} # aw, heck
In which case I gather `__mutable_args__` is superfluous, and the solution involves simply making sure that class instances avoid pointing to the same memory objects?

Thanks for this writeup.

By "what is your proposed solution?" I meant to ask "how would you change the language semantics to remove this 'idiotic' footgun", not "how would you work around this footgun?" I apologize for the ambiguity.

I initially thought this specific footgun was an unavoidable consequence in languages that primarily pass around references (like Python and Javascript). But it turns out that Javascript's default arguments do "make a special case exception for when the expression that constitutes a function default is evaluated!"


Yep, a renowned Python footgun. Although Lambda container reuse adds a modern twist to it.

Can I ask what editor you were using? Because decent ones should have warned you about mutable default args.

Shout out to the amazing Pyright, the fast and feature-complete python LSP that you can use with any editor that supports the LSP protocol!

I'm using VSCode with different Python specific extensions, like Pylance which uses Pyright, but it didn't warn me about the mutable function parameter. RIP

Huh, that sucks. I use Intellij Ultimate for pretty much everything, and it (and the Python focused Pycharm) will definitely warn on this particular Python quirk.

You're right, I checked, and pyright only warns about function calls in default arguments, not mutable containers.

I started a discussion in Pyright's repo here[0].

[0] https://github.com/microsoft/pyright/discussions/2306

Thanks to the maintainer's great responsiveness, Pyright will implement such a warning soon: https://github.com/microsoft/pyright/issues/2308

Linters can help a ton with stuff like this, I think pylint warns against this footgun.

Also is there a reason for not using the shorter:

var = var or []

Instead of:

var = [] if var is None else var

Both mypy and pyright/pylance do the right thing when things are guarded with explicit None checks and the wrong thing with “or”, so even though I prefer “or” otherwise when its gated upstrean so that None is the only falsey value it should get, I’ve taken recently to using explicit checks to make typecheckers happy.

That's not been my experience. I frequently use or way and have both mypy/pyright enabled. Testing this toy function,

from typing import List, Optional

def f(x: Optional[List[int]]): x = x or [] reveal_type(x)

mypy and pyright both show List[int] as type of x at the end and correctly drop None.

I've never thought of this as a footgun. It's useful to accumulate state in a function, sort of like a static variable in c. E.g.,

  def func(x=[]):
      return x
I've always been annoyed that pylint yells at you for this.

The issue is that it’s a non-apparent side effect to an outside caller.

I think using `or` for default fallbacks is generally considered bad practice because it disallows all falsey values, like `0` and `{}`.

It doesn't matter in this particular case because the only falsey List is `[]` anyway, but it's a bad habit.

JavaScript added the `??` operator for this specific problem.

As a Rubyist who only uses Python casually, it's neat to learn this difference! You can write very similar code in Ruby but such default objects appear to be created fresh on initialization avoiding the problem.

OTOH, Ruby has a very similar footgun with mutable defaults, just with default values of Hashes, not default method arguments.

Having a bug caused by mutable state occur inside a "lambda function" really throws into relief what a misnomer that is.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact