The cmp_to_key func acts as a class factory for generating key objects that behave according to a user defined cmp function. 

Many patterns/libs that make use of key functions (for example blist and the SortedCollection recipe) store the the results of the key function internally to provide faster subsequent operations. For large datasets, the number of key objects can become large, and it makes sense to keep their memory footprint as small as possible. 

From this stance, I propose to add "__slots__ = ('obj',)" to the class definition of "K" in cmp_to_key to prevent the creation of the un-needed instance dict and save the memory. 

I find it hard to conjure up a case where a user would need to weakref or add attributes to the return value of a key func. But should that need arise, that functionality could be provided as optional kwargs to cmp_to_key, or just have the user subclass the return value of cmp_to_key. Alternatively, the addition of __slots__ could be specified as a kwarg to cmp_to_key. 

I'd be happy to make a patch for any of these (or alternative) options.

I think this is a good idea.
Am not sure whether it should be backported though.

Right, I edited the versions field to match the versions that contain cmp_to_key, but if this behavior change is just for performance, it’s not a candidate for backport.

Chris: I’d go with the simpler implementation.

New changeset 97799a54728e by Raymond Hettinger in branch '3.2':
Issue #11628: cmp_to_key should use __slots__.
http://hg.python.org/cpython/rev/97799a54728e

New changeset 2e5e5d515c09 by Raymond Hettinger in branch 'default':
Issue #11628: cmp_to_key should use __slots__.
http://hg.python.org/cpython/rev/2e5e5d515c09

Man, you guys are quick! Cheers!

New changeset 136298326897 by Raymond Hettinger in branch '2.7':
Issue #11628: cmp_to_key should use__slots__.
http://hg.python.org/cpython/rev/136298326897