According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

Nope, it returns either Hashable or _HashedSeq inherited from list(not hashable!) because of:

>>> from functools import _make_key
>>> _make_key(('aa'),{},False)
['a', 'a']
>>> type(_make_key(('aa'),{},False))
<class 'functools._HashedSeq'>
>>> _make_key((1,),{'a':'b'},False )
[1, <object object at 0x7fbc462daeb0>, 'a', 'b']
>>> type(_make_key((1,),{'a':'b'},False ))
<class 'functools._HashedSeq'>

# if len(args) == 1 and type(key[0]) in fasttypes, it returns one of the fasttypes: {int,str} which is hashable

>>> _make_key(('a'),{},False)
'a'
>>> type(_make_key(('a'),{},False))
<class 'str'>
>>> _make_key(([1]),{},False)
1
>>> type(_make_key(([1]),{},False))
<class 'int'>

_HashedSeq is hashable, because it defines __hash__. That's the whole purpose of having _HashedSeq instead of just using a list.

Well, it turns out to be hashable, but I thought maybe it inherits from list would make it not Hashable. Then, I decided to test it myself.

>>> from collections.abc import Hashable 
>>> isinstance(_make_key((1,),{'a':'b'},False ), Hashable)
True

Nice catch!

Maybe the __repr__ made me think it is a list object.

At the end of the day, I think no one would like to annotate a variable like:

key: _HashedSeq =  _make_key(args, kwargs, False ) 

But rather:

key: Hashable =  _make_key(args, kwargs, False ) 

However, giving the programmer more options would be kind.

If you need the key to be a _HashedSeq specifically:

• Before this PR, you need assert isinstance(key, _HashedSeq) to convince mypy that it is actually a _HashedSeq. Otherwise mypy thinks "it can be _HashedSeq or any other Hashable, so let's error if you assume it's _HashedSeq".
• After this PR, the whole _HashedSeq is gone from the stubs and assert isinstace(key, _HashedSeq) will error. We can add it back later if someone actually wants to do this.

In general, we support undocumented non-public things (name starts with _, like _make_key and _HashedSeq) only if someone actually needs them, and if anyone needs this, we can add _HashedSeq back.