>>> isinstance({}.keys(), collections.Set)
True
>>> [method for method in set(dir(collections.Set)) - set(dir({}.keys()))
...  if not method.startswith('_')]
['isdisjoint']

(in Python 2.7, use "viewkeys()" instead of "keys")

dict_items has the same problem.

Concrete classes are allowed to have more features than the corresponding ABC. The ABCs are not intended to be full-featured; instead, they specify the part of the API that will be guaranteed. For example, the union() method for built-in sets allows two or more set arguments, but the corresponding method for the ABC is limited to two arguments. This was an intentional difference, designed to make it easier for people to implement a Set class.

That being said, the omission of isdisjoint() was an oversight and I see no reason that this shouldn't be fixed.

In this case, the concrete class is the one missing a method.

Concrete classes are allowed to provide more features than the corresponding ABC, but the converse is not true to the best of my knowledge.

dict_keys .register()s as supporting the Set ABC, so it does not automatically pick up the method through inheritance. Put another way:

>>> # dict_keys provides the Set ABC API
>>> isinstance({}.keys(), collections.Set)
True

>>> # The Set ABC provides isdisjoint
>>> hasattr(collections.Set, 'isdisjoint') 
True

>>> # Ergo, dict_keys should provide isdisjoint ... but it doesn't
>>> hasattr({}.keys(), 'isdisjoint')       
False

See also bpo-9213 for another case where a concrete class is missing a method provided by an ABC it claims to support.

I sort of wonder if .register() should verify that the concrete class provides all of the methods of the ABC.

I had misread you original post. Thought you we saying that the Set ABC was missing disjoint. Disregard my last post.

The attached patch adds the isdisjoint method to dict_keys and dict_items.
Pseudocode for the method:

def isdisjoint(self, other):
    if self is other:
        if len(self) == 0:
            return True
        else:
            return False
    else:
        for item in other:
            if item in self:
                return False
        return True

Titles that fit in the box, like 'Dict.keys lacks .isjoint method.' are easier to read and keep track of.

Unless there is a reason I have missed, I would iterate through the smaller set, which might even be empty or nearly so, rather than either in particular.

Unless there is a reason I have missed, I would iterate through the
smaller set, which might even be empty or nearly so, rather than
either in particular.

You're right, here is a new patch. Pseudocode:

def isdisjoint(self, other):
    if self is other:
        if len(self) == 0:
            return True
        else:
            return False
    else:
        if len(other) > len(self):
            self, other = other, self
        for item in other:
            if item in self:
                return False
        return True

Thank you for the patch.

We should only iterate over the shorter set if the longer set is really a set and not just a sequence. PySequence_Contains may take O(n) time on a list, making the algorithm an expensive O(n**2) overall. I note that set_isdisjoint doesn't try to examine the lengths.

Also, since PyIter_Next returns NULL to indicate the end of the iteration OR to indicate an exception, the end of the function should look like this:

    Py_DECREF(it);
    if (PyErr_Occurred())
        return NULL;
    Py_RETURN_TRUE;

Other than those two issues, the patch looks good to me.

Thanks for the corrections. I'm attaching the new patch as issue9212b.diff. I'm using PyAnySet_Check to determine if the other argument is really a set, but I'm not entirely sure, that this is correct. Please let me know if other corrections are needed.

It would be nice to get this fixed before the next release.

I will aim to spend some time with this (and the similar Issue bpo-9213) today and/or tomorrow, so that it can be committed in time for 3.2a2.

Committed to py3k in r84435. Raymond, do you want to look the commit over before I merge it into 3.1 and 2.7?

Meant to add:

I made some relatively minor changes to Daniel Urban's patch. Mostly, I rearranged the order of a few things to avoid unnecessary work (e.g., only compute "len_other" if we've already checked that "other" is a set).

Thank you Daniel for the patch! :-)

Also, credited Daniel Urban for the patch in r84436 (forgot that the first time around -- sorry!).