GeezerVison

Please fix bpo number

Hi @GPery, thanks for taking the time to improve Python!

As @tirkarthi mentionned, #11781 is already open to add this functionnality thought.

#11781 adds a key parameter, which I believe is inferior. It's a specific case of a comparison callback, and in my use case would force me to create another class to serve as a comparator for my objects, at which point I might as well wrap them and add lt.

Furthermore, I believe this is more in-line with similar standard functions in other languages such as C++ (std::sort), Java (PriorityQueue) or Rust (slice.sort_by).

key is more in line with Python list.sort() and sorted().
Use cmp_to_key() for handling __lt__ case.

@GPery, both are actually two ways to describe the same behavior, do you know functools.cmp_to_key (https://docs.python.org/3/library/functools.html#functools.cmp_to_key) to go from one to the other?

You're right, both cases are about the same.
The key extraction syntax still seems less intuitive to me than a compare function. Having been adopted by so many other languages, I think it should be the preferred behaviour, despite not matching list (after all, it's a separate library, even though it's builtin).

I feel it might be relevant to note that this isn't reverting to the old-style (C-style) comparison.
It's a boolean comparison, parallel to how __lt__ behaves.
Not the old style, but a different alternative to it.

list sort uses key because this is much faster than old cmp method (very similar to your current proposal).
The reason is: key is calculated only once per item but cmp is called for every compared pair.
This is doesn't matter for fast compiled languages like C++ or Rust but in Python the python function call is relatively expensive.

"every compared pair" is also (up to) once per item.
If we're on the subject of function calls, the key flow calls the key function, then the result's __lt__ function per item.
The compare_function flow only calls the callback.

The reason is: key is calculated only once per item but cmp is called for every compared pair.

What's your point exactly? Even with custom keys, you still need to compare every compared pair with __lt__.

(Note: I do think that key is the way to go but only for compatibility with sorted(), not for performance).

@jdemeyer IIRC at least for list sorting the point is that python-provided comparison function is usually much slower than built-in compare methods for int, str or tuple.

at least for list sorting the point is that python-provided comparison function is usually much slower than built-in compare methods for int, str or tuple.

OK, so the performance argument becomes "in certain use cases, using key would be faster than a Python 2 style cmp". Fair enough. But there are certainly also uses cases (arguably less common) where the opposite is true, for example when using cmp_to_key.

@asvetlov I'm not sure what were the original concerns, but I checked this PR and the key PR (#11781) with timeit, compare_function is extremely slightly, yet consistently, faster.

timeit -s "from bisect import bisect
                          class C:
                           def __init__(self, n):
                            self.n = n
                          data = [C(n) for n in range(1_000_000)]
                          cmp = C(25)" "bisect(data, cmp, key=lambda x: x.n)"
50000 loops, best of 5: 6.93 usec per loop

timeit -s "from bisect import bisect
                          class C:
                           def __init__(self, n):
                            self.n = n
                          data = [C(n) for n in range(1_000_000)]
                          cmp = C(25)" "bisect(data, cmp, compare_function=lambda a,b: a.n < b.n)"
50000 loops, best of 5: 6.79 usec per loop

I think this is a rather fair comparison.
Anyway, the real issue is syntax compatibility, which I think is subjective.

I just tried and had the opposite result, both with and without pydebug, I wonder what exactly might be doing that.

Hi there, why was this one closed? Was it superseded/WONTFIXED? This addition would be amazing, I keep hitting it. Let me know if there's any way I can help. Thanks to all.