REF: IntervalIndex.get_value partial slicing

[jbrockmendel]

In trying to refactor IntervalIndex.get_loc/get_value to more closely follow the other implementations, I stumbled on a behavior in a test case that surprised me:

idx = IntervalIndex.from_tuples([(1, 5), (3, 7)])
s = Series(range(len(idx)), index=idx)
key = [4]

>>> idx.get_loc(key)  # <-- behaves like all the other get_locs in requiring scalar
KeyError: [4]

>>> idx.get_value(s, key)   # <-- surprising
(1, 5]    0
(3, 7]    1
dtype: int64

All of the other FooIndex.get_value methods raise InvalidIndexError on non-scalar.

AFAICT this is behaving sort of like a slicing operation, since 4 is in the interior of both of these intervals. @jschendel is that correct? If so, we should try to refactor IntervalIndex.get_value to explicitly dispatch for this case, kind of like DatetimeIndex._get_string_slice

[jschendel]

This should follow the other get_value methods in that a non-scalar should raise an InvalidIndexError, or at least I don't see a reason why it shouldn't. The original implementation accepted non-scalars though, so it's possible there's some history there from before my time that I'm unaware of (cc @jreback).

I would expect the scalar 4 to return the same result as above for idx.get_value(s, 4), which it does on master. For IntervalIndex we want the indexing methods to work for exact Interval matches (i.e. same left/right/closed) or for points contained in the intervals. This does not apply to overlapping or nested Interval matches though, e.g. Interval(2, 6) or Interval(1.5, 2.5) should both raise a KeyError, which is a somewhat recent change (implemented during the sprint last summer).

[mroeschke]

Looks like this is consistent with get_loc now. Could use a test.

In [92]: idx = IntervalIndex.from_tuples([(1, 5), (3, 7)])
    ...: s = Series(range(len(idx)), index=idx)
    ...: key = [4]

In [93]: idx.get_value(s, key)
InvalidIndexError: [4]

In [94]: idx.get_loc(key)
InvalidIndexError: [4]

In [95]: pd.__version__
Out[95]: '1.1.0.dev0+1390.gf3fdab389'

[primaprashant]

take