Why is pd.Index.union not commutative?

[aa1371]

Code Sample, a copy-pastable example if possible

>>> ix = pd.Index([1,2])
>>> eix = pd.Index([])
>>> pi = pd.PeriodIndex(['19910905', '19910906'], freq='D')

# Pair 1
>>> pi.union(eix)
ValueError: can only call with other PeriodIndex-ed objects
>>> eix.union(pi)
PeriodIndex(['1991-09-05', '1991-09-06'], dtype='period[D]', freq='D')

# Pair 2
>>> pi.union(ix)
ValueError: can only call with other PeriodIndex-ed objects
>>> ix.union(pi)
Index([1, 2, 1991-09-05, 1991-09-06], dtype='object')

Problem description

Conceptually I would imagine a union operation to be commutative. I was just wondering if there was an deliberate rationale behind not implementing pd.Index._assert_can_do_setop to only fail if the complementary self._assert_can_do_setop also fails.

This behavior also leads to some unexpected behaviors in pd.concat. For example:

>>> df1 = df1 = pd.DataFrame([[1,2,3],[1,2,3]], index=pd.PeriodIndex(['19910905', '19910906'], freq='D'))
>>> df2 = pd.DataFrame()
>>> pd.concat([df1, df2], axis=1, keys=['a', 'b'])
ValueError: can only call with other PeriodIndex-ed objects
>>> pd.concat([df2, df1], axis=1, keys=['a', 'b'])
Works!

Additionally (and perhaps this should be raised as a separate issue) should the specific implementation of pd.PeriodIndex._assert_can_do_setop not raise if the other index is empty? Since pd.Index([]).union(<instance of pd.PeriodIndex>) results in an instance of pd.PeriodIndex.

[gfyoung]

cc @jreback

[jreback]

we don't ignore empties. actually these should all convert to object dtype (and work).

[TomAugspurger]

Agreed.

[aa1371]

@jreback - just to clarify - are you suggesting that:

1. the union of an empty index with any other index should not result in an index of the same type as the "other" index.
   e.g.

>>> pd.Index([]).union(pd.period_range('19910905', periods=2))
Current: PeriodIndex(['1991-09-05', '1991-09-06'], dtype='period[D]', freq='D')
Desired: Index([1991-09-05, 1991-09-06], dtype='object')
>>> pd.Index([]).union(pd.interval_range(start=0, end=5))
Current: IntervalIndex([(0, 1], (1, 2]]
                       closed='right',
                       dtype='interval[int64]')
Desired: Index([(0, 1], (1, 2]], dtype='object')

2. Any index should be able to form a union with any other type, as long as the resultant dtype is object?
   e.g.

>>> pd.period_range('19910905', periods=2).union(pd.Index([1,2,3]))
Current: ValueError
Desired: Index([1991-09-05, 1991-09-06, 1, 2, 3], dtype='object')

2b) And should this only be the case for which the _assert_can_do_setop is valid for at least one of the indexes?
e.g.
Consider the following pairs:
Index/PeriodIndex - Index can do setop with PeriodIndex, but not vice-versa.
IntervalIndex/PeriodIndex - neither index can do a setop with the other - should this still work and result in an object dtype?

[jreback]

yes to both

[aa1371]

Thoughts on 2b?

[jreback]

no

like indexes are combined to form the same type
unlike are coerced to object and then combine

it’s likely that the coercion logic is not fully handling things

we should not have any special cases - these are general ops

[aa1371]

@jreback - how should the union of Int64Index and RangeIndex behave ideally? Should they also result in an object dtype or behave the same as they do now?

[TomAugspurger]

RangeIndex should always be just an optimization of Int64Index. So you would return a RangeIndex if possible, else an Int64Index.

…

On Thu, Nov 8, 2018 at 9:03 PM ArtinSarraf ***@***.***> wrote: @jreback <https://github.com/jreback> - how should the union of Int64Index and RangeIndex behave ideally? Should they also result in an object dtype or behave the same as they do now? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#23525 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABQHIu9_YxMmMlL0xNVPSc523ZESwKX1ks5utPCKgaJpZM4YPp62> .