question: if there are two pushkins (say, for GCM and APNS), do they both correctly refer to the same instance of Counter ? or do they each get their own instance (which will probably confuse the prometheus lib)?

(I can never quite remember how class-level attributes work)

They should get the same one. I wasn't 100% sure that's what was wanted, but I think it is.

My understanding was/is that the class attribute is initialised when the class declaration is executed, and that attribute stays with the class.

Note it is always referred to as ConcurrencyLimitedPushkin.RATELIMITING_DROPPED_REQUESTS, so it would always act like a namespaced global.

I didn't know whether subclasses would get access to it (not that it matters, since we access the field statically every time and not through self or cls), so I tried it:

oli@bbm-neon:~$ python3
Python 3.6.9 (default, Jul 17 2020, 12:50:27) 
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> class X:
...     bob = []
... 
>>> class Y(X):
...     pass
... 
>>> class Z(X):
...     pass
... 
>>> id(Y.bob)
140481390301896
>>> id(X.bob)
140481390301896
>>> id(Z.bob)
140481390301896

(and to answer the question, yes this is intended — the pushkin is distinguished by the pushkin label, as this is what Prometheus understands)

In your example, even doing the following should return the same thing:

id(Y().bob)

Doesn't matter if you access it via the class name or not (although I think it is clearer to access class-level variables that way).

In your example, even doing the following should return the same thing:

id(Y().bob)

Doesn't matter if you access it via the class name or not (although I think it is clearer to access class-level variables that way).

Yes you are right (and I suspected so), but my experience from Java is that 'access static vars by class name, even if you can do it by instance' :)