Creating this issue by request of INADA Naoki to discuss my proposed patch in #5944.

Copy/pasting from that PR:

If you try something like issubclass('not a class', str), you get a helpful error message that immediately clues you in on what you did wrong:

>>> issubclass('not a class', str)
TypeError: issubclass() arg 1 must be a class
("AHA! I meant isinstance there. Thanks, friendly error message!")

But if you try this with some ABC, the error message is much less friendly!

>>> from some_library import SomeAbc
>>> issubclass('not a class', SomeAbc)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/Cellar/python3/3.6.4_2/Frameworks/Python.framework/Versions/3.6/lib/python3.6/abc.py", line 230, in __subclasscheck__
    cls._abc_negative_cache.add(subclass)
  File "/usr/local/Cellar/python3/3.6.4_2/Frameworks/Python.framework/Versions/3.6/lib/python3.6/_weakrefset.py", line 84, in add
    self.data.add(ref(item, self._remove))
TypeError: cannot create weak reference to 'str' object

("WTF just went wrong?" Several more minutes of head-scratching ensues. Maybe a less experienced Python programmer who hits this hasn't seen weakrefs before and gets overwhelmed, maybe needlessly proceeding down a deep rabbit hole before realizing no knowledge of weakrefs was required to understand what they did wrong.)

Or how about this example:

>>> from collections import Reversible
>>> issubclass([1, 2, 3], Reversible)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/Cellar/python3/3.6.4_2/Frameworks/Python.framework/Versions/3.6/lib/python3.6/abc.py", line 207, in __subclasscheck__
    ok = cls.__subclasshook__(subclass)
  File "/usr/local/Cellar/python3/3.6.4_2/Frameworks/Python.framework/Versions/3.6/lib/python3.6/_collections_abc.py", line 305, in __subclasshook__
    return _check_methods(C, "__reversed__", "__iter__")
  File "/usr/local/Cellar/python3/3.6.4_2/Frameworks/Python.framework/Versions/3.6/lib/python3.6/_collections_abc.py", line 73, in _check_methods
    mro = C.__mro__
AttributeError: 'list' object has no attribute '__mro__'
Here you don't even get the same type of error (AttributeError rather than TypeError), which seems unintentionally inconsistent.

This trivial patch fixes this, and will hopefully save untold numbers of future Python programmers some time and headache.

Let me know if any further changes are required, and thanks in advance for reviewing.

Why issubclass() doesn't check it? Maybe, non-type class is supported by Python. But I'm not sure. I'm not meta programming expert.

But we use "duck typing". In this case, if the object (a) supports weakref and (2) has __mro__ and it is tuple, we treat the object as a class.

And when raising TypeError, information about which assumption was broken may be useful.

So I'm -1 on adding such strong check or removing internal information without meta programming expert advice.

ABC.register() has an explicit check, and it is mentioned in PEP-3119. The point here is not to change issubclass(), but to change ABC.__subclasscheck__(). It may conceivably have stricter requirements than issubclass() has. But certainly an advice from actual ABC users would be nice.

https://bugs.python.org/msg313396

issubclass(class-like, class-like) is allowed.
I don't think raising type error for issubclass(class-like, ABC) is good idea. It should return False.

Hmm, normal class doesn't support issubclass(class-like. class).

Python 3.8.0a0 (heads/master:fc7df0e664, Mar  8 2018, 09:00:43)
[GCC 7.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import typing
>>> issubclass(typing.MutableMapping, object)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: issubclass() arg 1 must be a class
>>> issubclass(typing.MutableMapping, type)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: issubclass() arg 1 must be a class
>>> issubclass(typing.MutableMapping, typing.Mapping)
True

OK, problem is ABC should support it or not.

I do not see any point in allowing non-types in ABCMeta.__subclasscheck__. Currently, ABCs are clearly not designed to support non-types:

1. ABCMeta.register() accepts types only.
2. ABCMeta.__subclasscheck__ implicitly requires its arguments to support weak references (regardless of whether __subclasshook__ is called or not). This requirement alone doesn't make sense, so it seems to be an exposed implementation detail stemming from the fact that non-types were not intended to be supported.
3. Some ABC users already expect that the argument of __subclasshook__ is a type (see the example with collections.abc.Reversible by OP).
4. Attempting to support arbitrary arguments in ABC.__subclasscheck__ (by returning False instead of raising TypeError or worse) will not solve any 'issubclass' inconsistencies. 'issubclass' is fundamentally "fragmented": issubclass(x, y) may return True/False while issubclass(x, z) may raise TypeError, depending on __subclasscheck__ implementation. It may be too late to impose stricter requirements for the first argument of issubclass because 'typing' module relies on the support of non-types there.

1. ABCMeta.register() accepts types only.

Yes. While ABC.register() and issubclass() have different users (e.g. ABC.register() will be used by framework author, and issubclass will be used by framework users), it's positive reason to remove non-type support.

2. ABCMeta.__subclasscheck__ implicitly requires its arguments to support weak references (regardless of whether __subclasshook__ is called or not). This requirement alone doesn't make sense, so it seems to be an exposed implementation detail stemming from the fact that non-types were not intended to be supported.

Isn't it just a limitation?
Most Python-implemented objects supports weakref. I don't think "requiring weakref support implies it must be type object".

3. Some ABC users already expect that the argument of __subclasshook__ is a type (see the example with collections.abc.Reversible by OP).

What "by OP" means?
I can't find if not issubclass(cls, type): raise TypeError in Reversible implementation.
They do duck-typing, same to ABC.

4. Attempting to support arbitrary arguments in ABC.__subclasscheck__ (by returning False instead of raising TypeError or worse) will not solve any 'issubclass' inconsistencies. 'issubclass' is fundamentally "fragmented": issubclass(x, y) may return True/False while issubclass(x, z) may raise TypeError, depending on __subclasscheck__ implementation.

Yes, as I commented above.

It may be too late to impose stricter requirements for the first argument of issubclass because 'typing' module relies on the support of non-types there.

Of course.

Personally speaking, I dislike magics including ABC, __subclasscheck__, overwriting __class__ with dummy object. So I'm OK to limit the ability of it.

But I don't know much about how mages use ABC. I need mages comment before merging the pull request.

BTW, do you think it should be backported to 3.7, or even 3.6?
Can fc7df0e be reverted?

Isn't it just a limitation?
Most Python-implemented objects supports weakref. I don't think "requiring weakref support implies it must be type object".

Formally, there is no implication. It is the abc module authors who know the truth. But I can't imagine why anybody would impose such a limitation by design, because while instances of user-defined classes support weakrefs, built-in classes used by everybody like tuple, list and dict don't. That's why I guessed that non-types were not meant to be supported.

What "by OP" means?
OP = Original poster (@jab).

I can't find if not issubclass(cls, type): raise TypeError in Reversible implementation.
They do duck-typing, same to ABC.

Sorry for being unclear. There is no explicit check as you say, but __mro__ is directly accessed (see msg313376). But it may probably be considered "duck typing" too.

But I don't know much about how mages use ABC. I need mages comment before merging the pull request.
Totally agree.

BTW, do you think it should be backported to 3.7, or even 3.6?
3.7 certainly has my vote -- this can hardly be considered a new feature.

For 3.6, I'd listen to ABC users/experts. Might raising a TypeError instead of returning False from issubclass(user_defined_obj, ABC) break something important? Personally, I think it would mostly expose bugs and not hinder reasonable usage.

Can fc7df0e be reverted?

Seems like it can, but the test should survive in some form :)

Can fc7df0e be reverted?

Even if subclass() will check explicitly that its first argument is a type, ABC.__subclasscheck__() can be called directly, and it shouldn't crash when called with non-type.

PR 5944 changes ABC.__subclasscheck__ (not issubclass) to check its first argument, so if it's merged there will be no crash even with the revert.

> Isn't it just a limitation?
> Most Python-implemented objects supports weakref. I don't think "requiring weakref support implies it must be type object".

Formally, there is no implication. It is the abc module authors who know the truth. But I can't imagine why anybody would impose such a limitation by design, because while instances of user-defined classes support weakrefs, built-in classes used by everybody like tuple, list and dict don't. That's why I guessed that non-types were not meant to be supported.

Of course, issubclass(42, AnyABC) must raise TypeError. They aren't class-like object. I didn't discuss on it.

I talked about class-like objects.
For example, this code works on Python 3.6, but not on 3.7. typing.Mapping looks like type, but it is just an instance for 3.7.

  import typing
  import collections.abc as cabc
  print(issubclass(typing.MutableMapping, cabc.Mapping))  # Python 3.7 raises TypeError

I don't think it's real problem. But if someone claims it's real issue, we can make typing.MutableMapping more "class-like" by adding __mro__.

diff --git a/Lib/typing.py b/Lib/typing.py
index 7ca080402e..2edaa3f868 100644
--- a/Lib/typing.py
+++ b/Lib/typing.py
@@ -619,6 +619,7 @@ def __init__(self, origin, params, *, inst=True, special=False, name=None):
                               a for a in params)
         self.__parameters__ = _collect_type_vars(params)
         self.__slots__ = None  # This is not documented.
+        self.__mro__ = (origin,) + getattr(origin, '__mro__', (object,))
         if not name:
             self.__module__ = origin.__module__

Again, I don't think it's a real problem.
Maybe, we can add the check, and revert it if someone claims.

To be honest I am still undecided on this. In principle, I am OK with status quo, but I am also OK, with the PR that will prohibit non-classes. I am a bit worried that it may break some existing code, so it is probably not for 3.7.

Regarding status quo (expanding the examples of @inada.naoki and @jab):

>>> import typing
>>> import collections.abc as cabc
>>> issubclass(typing.Mapping, cabc.Mapping)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/izbyshev/workspace/cpython/Lib/abc.py", line 143, in __subclasscheck__
    return _abc_subclasscheck(cls, subclass)
TypeError: issubclass() arg 1 must be a class
>>> from abc import ABC
>>> issubclass(typing.Mapping, ABC)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/izbyshev/workspace/cpython/Lib/abc.py", line 143, in __subclasscheck__
    return _abc_subclasscheck(cls, subclass)
  File "/home/izbyshev/workspace/cpython/Lib/abc.py", line 143, in __subclasscheck__
    return _abc_subclasscheck(cls, subclass)
  File "/home/izbyshev/workspace/cpython/Lib/contextlib.py", line 30, in __subclasshook__
    return _collections_abc._check_methods(C, "__enter__", "__exit__")
  File "/home/izbyshev/workspace/cpython/Lib/_collections_abc.py", line 73, in _check_methods
    mro = C.__mro__
  File "/home/izbyshev/workspace/cpython/Lib/typing.py", line 706, in __getattr__
    raise AttributeError(attr)
AttributeError: __mro__
>>> ABC.register(int)
<class 'int'>
>>> issubclass(typing.Mapping, ABC)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/izbyshev/workspace/cpython/Lib/abc.py", line 143, in __subclasscheck__
    return _abc_subclasscheck(cls, subclass)
TypeError: issubclass() arg 1 must be a class
>>> typing.Mapping.__mro__ = ()
>>> issubclass(typing.Mapping, ABC)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/izbyshev/workspace/cpython/Lib/abc.py", line 143, in __subclasscheck__
    return _abc_subclasscheck(cls, subclass)
TypeError: issubclass() arg 1 must be a class
>>> typing.Mapping.__bases__ = ()
>>> issubclass(typing.Mapping, ABC)
False

Can't say that I'm OK with it :)

I'm for forbidding non-types in ABCMeta.__subclasscheck__, but if we are to add clean support for "class-likes" instead, I think that "class-like" objects should be clearly defined, for example, that they must have __mro__ and __bases__ (and probably support weakrefs unless we want to skip caching if they don't). ABCMeta.__subclasscheck__ is not standalone: it relies on issubclass() and __subclasshook__, and both of them have some expectations in practice.

My current opinion is:

• -1 for 3.6: Behavior should not be changed without strong reason, even the behavior is not documented.
• +1 for 3.8: I like strict and less magic.
• +0 for 3.7: beta3 is bit late, but this change has very little chance to cause real world problem.

I agree except that I'd like to see it in 3.7 too.

I'll share the use case that prompted me to submit this PR in the first place.

I am the author of bidict (https://pypi.python.org/pypi/bidict), which provides a bidirectional dict class. A bidict is just like a dict, except it automatically maintains its inverse bidict, which is accessible via its .inv attribute. To prevent a bidict and its inverse from creating a strong reference cycle, a weak ref is used to store the reference one direction.

bidicts implement my BidirectionalMapping ABC, which extends collections.abc.Mapping to include the .inv abstractproperty. BidirectionalMapping overrides __subclasshook__ so that outside implementations that don't subclass it explicitly may still be considered subclasses.

Recently, I tried something like issublass('foo', BidirectionalMapping), and got the "cannot create weak reference to 'str' object" error. Because this error message differs from the (much more helpful) "arg 1 must be a class" error message that you get when you do e.g. issubclass('foo', Mapping), I thought there might be a bug somewhere in my code. Then I looked deeper and found where this is really coming from.

I experimented more and noticed that issubclass('foo', Reversible) raises AttributeError, which isn't even the same type of error.

The exceptions that are raised in these cases seem like an abstraction leak. The error messages do not help users immediately realize what they did wrong and how they can fix it; more knowledge of internals is required to make sense of what's going on than should be needed. The inconsistency in these errors is a further problem. The same mistake should not cause three different errors unless there is some really good reason. This seems unintentional. Can any of the original authors say whether this is working as intended or if this is in fact an oversight?

The current behavior causes confusion for both less experienced and more experienced Python users alike. (Would anyone else here have correctly predicted all of the different errors that the examples above cause? How many other Python experts could have?) For less experienced users, Python's general consistency and predictability, lack of gotchas, and good errors are some of its best features. This is such an exception that it seems like a bug.

I'm happy for some other patch than the one I submitted in #5944 to land if necessary, as long as something fixes this. And fwiw, +1 for 3.7, unless anyone can demonstrate any credible risk.

Thanks for your consideration :)

I am still -1 on changing this in Python 3.7, unless Guido wants this in 3.7, if yes, then we can go ahead. Otherwise, I think we can consider just merging this into master, in this case I would switch to the PR to discuss the details.

While we're in feature freeze mode, we're not in release candidate mode
yet. I presume the code that would be patched in 3.7 is no different from
master? What's the concern? It doesn't seem anyone thinks it's risky?

If there are some code which depend on ABC.__subclasscheck__() allow class-like object, we'll have to revert it in 3.7b4 or rc.
I think it's the only risk.

Hmm... That is actually a not entirely imaginary risk, right? Can you come
up with a test program that would fail that way?

OTOH if we don't do this now, it's not going to be any easier to make this
change in 3.8. Maybe now's the time to experiment with it, and we can drop
it in rc1 if it causes problems. @ivan, your thoughts? Would you merge this
into master?

Would you merge this into master?

OK, I played with this a bit and it looks good. There is however a merge conflict now, and a NEWS item is missing. I will leave a comment in the PR.

New changeset 40472dd by Ivan Levkivskyi (jab) in branch 'master':
bpo-33018: Improve issubclass() error checking and message. (GH-5944)
40472dd

New changeset 346964b by Miss Islington (bot) in branch '3.7':
bpo-33018: Improve issubclass() error checking and message. (GH-5944)
346964b

New changeset f757b72 by INADA Naoki in branch 'master':
bpo-32999: Revert #50252 (fc7df0e) (GH-6189)
f757b72

New changeset 5d8bb5d by Ivan Levkivskyi (Miss Islington (bot)) in branch '3.7':
bpo-32999: Revert #50252 (fc7df0e) (GH-6189) (GH-6190)
5d8bb5d

I am closing this for now. We can re-open it later if problems will appear in 3.7.