On Friday, March 1, 2019 at 5:47:06 AM UTC-5, Steven D'Aprano wrote:On Fri, Mar 01, 2019 at 08:47:36AM +0200, Serhiy Storchaka wrote:
> Currently Counter += dict works and Counter + dict is an error. With
> this change Counter + dict will return a value, but it will be different
> from the result of the += operator.
That's how list.__iadd__ works too: ListSubclass + list will return a
value, but it might not be the same as += since that operates in place
and uses a different dunder method.
Why is it a problem for dicts but not a problem for lists?
> Also, if the custom dict subclass implemented the plus operator with
> different semantic which supports the addition with a dict, this change
> will break it, because dict + CustomDict will call dict.__add__ instead
> of CustomDict.__radd__.
That's not how operators work in Python or at least that's not how they
worked the last time I looked: if the behaviour has changed without
discussion, that's a breaking change that should be reverted.
Obviously I can't show this with dicts, but here it is with lists:
py> class MyList(list):
... def __radd__(self, other):
... print("called subclass first")
... return "Something"
...
py> [1, 2, 3] + MyList()
called subclass first
'Something'
This is normal, standard behaviour for Python operators: if the right
operand is a subclass of the left operand, the reflected method __r*__
is called first.
> Adding support of new operators to builting
> types is dangerous.
Explain what makes new operators more dangerous than old operators
please.
> > I do not understand why we discuss a new syntax for dict merging if we
> > already have a syntax for dict merging: {**d1, **d2} (which works with
> > *all* mappings). Is not this contradicts the Zen?
> >
> >
> >But (as someone else pointed out) {**d1, **d2} always returns a dict,
> >not the type of d1 and d2.
>
> And this saves us from the hard problem of creating a mapping of the
> same type.
What's wrong with doing this?
new = type(self)()
Or the equivalent from C code. If that doesn't work, surely that's the
fault of the subclass, the subclass is broken, and it will raise an
exception.
I don't think it is our responsibility to do anything more than call
the subclass constructor. If that's broken, then so be it.
Possibly relevant: I've always been frustrated and annoyed at classes
that hardcode their own type into methods. E.g. something like:
class X:
def spam(self, arg):
return X(eggs)
# Wrong! Bad! Please use type(self) instead.
That means that each subclass has to override every method:
class MySubclass(X):
def spam(self, arg):
# Do nothing except change the type returned.
return type(self)( super().spam(arg) )
This gets really annoying really quickly. Try subclassing int, for
example, where you have to override something like 30+ methods and do
nothing but wrap calls to super.
I agree with you here. You might want to start a different thread with this idea and possibly come up with a PEP. There might be some pushback for efficiency's sake, so you might have to reel in your proposal to collections.abc mixin methods and UserDict methods.
Regarding the proposal, I agree with the reasoning put forward by Guido and I like it. I think there should be:
* d1 + d2
* d1 += d2
* d1 - d2
* d1 -= d2
which are roughly (ignoring steve's point about types)
* {**d1, **d2}
* d1.update(d2)
* {k: v for k, v in d1.items() if k not in d2}
* for k in list(d1): if k not in d2: del d1[k]
Seeing this like this, there should be no confusion about what the operators do.
I understand the points people made about the Zen of Python. However, I think that just like with lists, we tend to use l1+l2 when combining lists and [*l1, x, *l2, y] when combining lists and elements. Similarly, I think {**d1, **d2} should only be written when there are also key value pairs, like {**d1, k: v, **d2, k2: v2}.
Best,
Neil
--
Steven
_______________________________________________
Python-ideas mailing list
Python...@python.org
https://mail.python.org/mailman/listinfo/python-ideas
Code of Conduct: http://python.org/psf/codeofconduct/