On Sat, Oct 04, 2014 at 10:43:02AM +0000, Antoine Pitrou wrote:

> Just because something is discussed doesn't mean it needs a PEP.

I know that in general discussions do not need a PEP, but we seem to be 
rushing awfully fast into writing code before we even know what the code 
is supposed to do. Perhaps "need a PEP" was too strong, but we surely 
need to do something to decide on constistent and well-defined 
behaviour. As you pointed out earlier, Counters aren't precisely 
multisets, they're more like a hybrid mapping/multiset. It's not clear 
what behaviour subset and superset should have for such a hybrid. By 
rushing into code before deciding on what the code is supposed to do, we 
end up with inconsistent behaviour.

Here are a pair of Counters which compare "is subset", but not "is 
subset or equal" using Ram's patch:

py> Counter(a=1, b=0) < Counter(a=2)
True
py> Counter(a=1, b=0) <= Counter(a=2)
False

On the other hand, here's a pair which compares "is subset or equal", 
but not ("is subset" or "equal"):

py> Counter(a=0) <= Counter(b=0)
True
py> (Counter(a=0) < Counter(b=0)) or (Counter(a=0) == Counter(b=0))
False

Adding elements with count=0 changes whether a set is a subset or not:

py> Counter(a=0) < Counter(b=0)
False
py> Counter(a=0) < Counter(b=0, c=0)
True
py> Counter(a=0, b=0) < Counter(b=0, c=0)
False

I'm not convinced that mixed Counter and regular dict comparisons should 
be supported, but Ram's patch supports any Mapping type. Is that a good 
idea? There's been no discussion on that question.

Can somebody explain to me what this means? How is this meaningfully a 
subset operation?

py> Counter(a="eggs") < Counter(a="spam")
True

I supported adding these operations to Counter, but the last thing I 
want is to enshrine a set of ad-hoc and inconsistent semantics that 
don't match standard multiset behaviour and are defined by the 
implementation, instead of having the implementation defined by the 
desired semantics.