This might of been suggested in the other thread, but,

Instead of going after the built-in set why not propose a collections based OrderedSet akin to OrderedDict? I would believe that might be less contentious.

On 24/08/2020 23:43, Cade Brown wrote:
Hello all,

I have a suggestion/idea for the Python standard (and/or a CPython implementation detail, considering its memory impact): having sets behave similar to dictionaries in that they preserve first-appearance order (i.e. dictionaries are ordered by key which was first inserted).

As is said in the Python standard, a 'set' is by definition unordered. Yet, internally they must be stored in some order, and when iterated upon, converted to string, etc, some ordering must be used. Right now, it is unspecified, which can lead to interesting results I noticed:

>>> print ({1, 2, 3}, {3, 2, 1})
{1, 2, 3} {1, 2, 3}
>>> print ({1, 5, 9}, {9, 5, 1})
{1, 5, 9} {9, 5, 1}

While, obviously, things like `A == B` are not affected by this, many algorithms may affect their output if sets are treated as iterables (again, the argument could be made, as it was for 'dict' as well, that they should not treat the order as relevant at all). Further, I believe that having a specification regarding the order and some guarantees could be useful and important to cohesiveness of Python as a whole.

Despite the obvious non-ordered nature of sets, I still think it is worth making them behave like `dict` objects (which are also, in some sense, 'unordered' by nature, and yet Python still specifies the order for 3.7+).

I would like to suggests that `set` objects are ordered by insertion, so that:
  * Sets have a defined, repeatable order for maximum reproducibility (assuming the code generates the set in a stable way)
  * Tests which are outside of Python can do string comparisons and expect the same output each time (as with `dict` objects generated in a predictable way).
  * On a high-level note, Python specifies more exact behavior. It is my belief (and I'm sure many others share this as well) that unspecified/implementation-dependent/otherwise-undependable features should be eliminated and replaced with exact semantics that do not suprise users.



Python-ideas mailing list --
To unsubscribe send an email to
Message archived at
Code of Conduct: