class C:
    def __init__(self, cond):
        if cond:
            self.a = 1
        self.b = 2

c1 = C(True)
c2 = C(False)

   
In Python 3.5, the dictionary keys are shared
---------------------------------------------

>>> sys.getsizeof(c2)
56
>>> sys.getsizeof(c1.__dict__)
96
>>> sys.getsizeof(c2.__dict__)
96

In Python 3.9, the keys are not shared
--------------------------------------

>>> sys.getsizeof(c2)
48
>>> sys.getsizeof(c1.__dict__)
272
>>> sys.getsizeof(c2.__dict__)
232

This represents an increase of memory use for c1 of 110%. (48+272)/(56+96) == 2.1

With compact object layout (https://github.com/faster-cpython/ideas/issues/69), 
any failure to share keys will result in a tripling of memory use per object.

It is not an uncommon pattern for some attributes to initialized lazily,
which causes also prevents key-sharing.

Not only does it increase memory use, it prevents optimizations that attempt to specialize attribute access for instances of a class, as different instances will have different layouts.

The purpose of compact dicts was to save memory, but in this case we are using a lot more memory by preventing PEP 412 from working.

ISTM this was mostly a net win.  We get compactness and sharing most of the time, and lose sharing only in cases like this where different instances of the same class happen to have different attributes.  Personally, I would not have expected sharing to occur in those cases.

Raymond,
When you say "this was mostly a net win" do you mean the more compact layout or ordering?

The compact layout is obviously a win, and doesn't conflict with sharing. The problem is that ordering conflicts with sharing.

As long as instances all have attributes that are a subset of the shared keys stored on the class, they can share attributes, or least they could in 3.5.

Duplicate of https://bugs.python.org/issue40116