The pprint pretty printer in Python 3 sorts sets/frozensets only if their length don't fit in one single line/row for the given width, else it was just leaving repr(my_set_instance) alone, like:

>>> import string, pprint
>>> pprint.pprint(set(string.digits))
{'7', '5', '2', '4', '1', '9', '6', '3', '0', '8'}

That order is quite random in Python 3.2+. But on Python 2.6 and 2.7, the result is shown as:
set(['0', '1', '2', '3', '4', '5', '6', '7', '8', '9'])

So for using pprint in doctests (or anything alike) with sets/frozensets, the pretty printer isn't as useful in Python 3 than it is in Python 2. The pprint tests for non-nested set/frozenset were only using some small ranges for testing. I've written a patch to solve that.

+1

It could reasonably be argued that not sorting is a bug for already-released 3.x versions.

+1 for treating this as a bug fix.

Since pprint() now supports all standard collections, I think it is worth to support them in saferepr() too. I have written an implementation and am finishing writing tests.

Wouldn't a fix for all standard collections be a fix for Python 3.5+, therefore another issue? http://bugs.python.org/issue23870

This issue is about sets/frozensets Python 3.2+, and I'm pretty sure it's backwards compatible, as I don't think any code running on Python 3.2.6 would depend on pprint randomness (how could?). Also, a multiline pprint would sort (tested with Python 3.2.6):

>>> pprint.pprint(set(string.digits), width=7)
{'0',
 '1',
 '2',
 '3',
 '4',
 '5',
 '6',
 '7',
 '8',
 '9'}

I see no reason to see a fix to this inconsistent behavior (sorting on multiline, not sorting on single line) as an enhancement just for a new Python 3.6 version. Besides being backwards compatible, the test_pprint was really verifying the order on set(range(n)) for small n, something that is already sorted by set.__repr__ but appears in test_pprint, which make me think it was intended as a pretty printer test, not as a set.__repr__ test.

Following patch adds support of all standard collections in pprint.saferepr().

There seems to be consensus that this should be treated as a bug fix, not a new feature. Could this still make it into 3.6 even though it missed the first beta?

Serhiy, would you be interested in converting your patch to a PR?