API: series_with_int64index[i:j] should be label-based

[jbrockmendel]

When we have a Int64Index, series[i] is treated as label-based. For consistency, series[i:j] should be too.

ser = pd.Series(range(5), index=pd.Index([1, 3, 5, 7, 9]))
ser2 = ser.copy()
ser2.index = ser2.index.astype('f8')

>>> ser[1]  # <- label-based
0
>>> ser[1:3]  # <- positional
3    1
5    2
dtype: int64

>>> ser2[1:3]  # <- label-based
1.0    0
3.0    1
dtype: int64

The logic by which ser[i:j] is positional is not easy to track down and does not use modern idioms (i.e. _should_fallback_to_positional)

NumericIndex._convert_slice_indexer specifically special-cases floating dtypes

    def _convert_slice_indexer(self, key: slice, kind: str):
        if is_float_dtype(self.dtype):
            assert kind in ["loc", "getitem"]

            # We always treat __getitem__ slicing as label-based
            # translate to locations
            return self.slice_indexer(key.start, key.stop, key.step)

        return super()._convert_slice_indexer(key, kind=kind)

while the base class ...

    def _convert_slice_indexer(self, key: slice, kind: str_t):
        # potentially cast the bounds to integers
        start, stop, step = key.start, key.stop, key.step

        # figure out if this is a positional indexer
        def is_int(v):
            return v is None or is_integer(v)

        is_index_slice = is_int(start) and is_int(stop) and is_int(step)
        is_positional = is_index_slice and not (self.is_integer() or self.is_categorical())

        if kind == "getitem":
            """
            called from the getitem slicers, validate that we are in fact
            integers
            """
            if self.is_integer() or is_index_slice:
                self._validate_indexer("slice", key.start, "getitem")
                self._validate_indexer("slice", key.stop, "getitem")
                self._validate_indexer("slice", key.step, "getitem")
                return key

AFAICT the is_positional = is_index_slice and not (self.is_integer() or self.is_categorical()) check should be is_positional = is_index_slice and self._should_fallback_to_positional. The if self.is_integer() or is_index_slice: should probably also be done in terms of _should_fallback_to_positional

Expected Behavior

>>> ser[1:3]  # <- match ser.loc[1:3]
1    0
3    1
dtype: int64

[prakhar987]

Updating is_positional with is_positional = is_index_slice and self._should_fallback_to_positional seems to work well. Though I can't fix if self.is_integer() or is_index_slice: without breaking some tests.

Will try to find out more.

[jbrockmendel]

Same underlying issue as #7501 IIUC