IntervalDtype inconsistencies and bugs

[jschendel]

1. Inconsistent comparisons versus string 'interval':

In [2]: IntervalDtype() == 'interval'
Out[2]: True

In [3]: IntervalDtype('interval') == 'interval'
Out[3]: False

In [4]: IntervalDtype('int64') == 'interval'
Out[4]: False

I'd expect all of these to return True, like how CategoricalDtype(*, *) == 'category' always returns True.

2. Inconsistent comparisons versus IntervalDtype(None):

In [5]: IntervalDtype(None) == IntervalDtype('interval')
Out[5]: False

In [6]: IntervalDtype(None) == IntervalDtype('int64')
Out[6]: False

I'd expect all of these to return True, like how CDT(None, None) == CDT(*, *) always returns True.

3. IntervalDtype.name attribute changes

In [7]: IntervalDtype().name
Out[7]: 'interval'

In [8]: IntervalDtype('interval').name
Out[8]: 'interval[]'

In [9]: IntervalDtype('int64').name
Out[9]: 'interval[int64]'

CategoricalDtype.name attribute is always the same:

In [10]: CategoricalDtype(list('abc'), True).name
Out[10]: 'category'

In [11]: CategoricalDtype(list('wxyz'), False).name
Out[11]: 'category'

I'd expect IntervalDtype.name to always return 'interval', like how CDT.name always returns 'category'. This makes the code for checking equality against strings (i.e. what I described in 1) simpler. I don't think the behavior of str(IntervalDtype) should change, which is currently the same as IntervalDtype.name, so I'd still have that return strings specifying the subtype.

4. CategoricalDtype gets cached incorrectly: (No longer an issue due to #19022)

In [12]: idt1 = IntervalDtype(CategoricalDtype(list('abc'), True))

In [13]: idt2 = IntervalDtype(CategoricalDtype(list('wxyz'), False))

In [14]: idt2.subtype
Out[14]: CategoricalDtype(categories=['a', 'b', 'c'], ordered=True)

This looks to be caused by the caching being done by string representation, and str(CDT(*, *)) always returns 'category':

pandas/pandas/core/dtypes/dtypes.py

Lines 673 to 679 in e1d5a27

	try:
	return cls._cache[str(subtype)]
	except KeyError:
	u = object.__new__(cls)
	u.subtype = subtype
	cls._cache[str(subtype)] = u
	return u

Can caching be removed entirely for IntervalDtype, or is there some need/advantage that I'm not seeing? Looking at the other dtypes, CategoricalDtype appears to have had the caching code removed, but PeriodDtype and DatetimeTZDtype are using it.

[jreback]

IIRC the reason for caching was:

a) perf (constructing a dtype each time involves some checks)
b) equality/hashing; the caching gives exactly the same object back so we can do is comparison. this is quite important as these are singular objects (though they are parameterized). IOW, IntervalDtype() should always be a singleton, and so on.