It would be nice to make random.sample and random.choice both have the same behavior with iterators. Currently random.sample accepts them happily, and whereas random.choice does not.  E.g.

> import random
> d = {'a':1, 'b':2}
> random.sample(d.keys(),1)
Out: ['a']

> random.choice(d.keys())
Out: TypeError: 'dict_keys' object is not subscriptable

random.choice could be redefined as follows to harmonize behavior, but I think the solution for random.choices maybe be more involved:

def choice(x): 
    random.sample(x,1)[0]

Seems similar to https://bugs.python.org/issue37682

The sample() function accepts d.keys() because a keys view is registered as being an instance of collections.abc.Set.  We have to do that because sample() is specifically documented as accepting sets.  If we had it do over again, that functionality would likely not be supported (it implicitly the population into a tuple).

For a fine-grained function like choice() that implicit population conversion would make no sense at all.  For choices(), we can choose not to make that API mistake again.

Thanks for the suggestion, but I am going to decline.  It isn't something that we need in practice and it would add complication for near zero gain.  (Also, there is no rule that all of these functions have to accept the same kinds of inputs -- different algorithms and different use cases allow for some flexibility).