Index.unique() should always return an Index object of the same type

[shoyer]

This should also be noted in the docstring for the method.

Currently, it sometimes returns numpy arrays:

>>> pd.DatetimeIndex(['2000-01-01', '2000-01-02']).unique()
DatetimeIndex(['2000-01-01', '2000-01-02'], dtype='datetime64[ns]', freq=None)

>>> pd.Index([0, 1, 2]).unique()
array([0, 1, 2])

Most of the work here is probably writing comprehensive tests to check each index type.

xref: https://github.com/pydata/pandas/pull/13361/files/17209f92330c5e949934aec9dea039b35faf6e40#r66179418

[jorisvandenbossche]

At the moment, I think DatetimeIndex is rather the exception, as most seem to return a numpy array (and CategoricalIndex a Categorical):

In [18]: pd.Index(['a', 'a', 'b']).unique()
Out[18]: array(['a', 'b'], dtype=object)

In [19]: pd.Int64Index([1,1,2]).unique()
Out[19]: array([1, 2], dtype=int64)

In [20]: pd.Float64Index([1,1,2]).unique()
Out[20]: array([ 1.,  2.])

In [21]: pd.CategoricalIndex(['a','a','b']).unique()
Out[21]:
[a, b]
Categories (2, object): [a, b]

[jreback]

this is a dupe of #4126

[jreback]

closing the other one actually.