REF: simplify json get_values

[jbrockmendel]

cc @WillAyd trying to get away from _internal_get_values in the json code; this is zero-logic-changing to narrow the scope of what gets passed to _internal_get_values.

AFAICT replacing the _internal_get_values call with calling __array__ doesnt break any tests, but in general the two are not always equivalent and it isnt obvious to me whether there are untested cases where it matters.

[jreback]

sure

[jbrockmendel]

@WillAyd could use your input on how to handle the last few edge cases. The relevant cases all involve either CategoricalIndex[dt64tz] or Series[Category[dt64tz]]. Example:

dti = pd.date_range('2016-01-01', periods=2, tz='US/Pacific')
ci = pd.CategoricalIndex(dti)
ser = pd.Series(dti)
cser = pd.Series(ci)
ciser = pd.Series(range(2), index=ci)

cser.to_json() matches ser.to_json(), but they get there in different ways. get_values in objToJSON calls ser._internal_get_values() and gets a ndarray[dt64_naive], while cser._internal_get_values() gets a ndarray[object] of Timestamp objects. Is going through object-dtype intentional, or would it be OK to match the non-categorical behavior?

If matching the non-categorical behavior is OK, then we can get equivalent behavior by calling ser.__array__ instead of ser._internal_get_values.

ciser.to_json currently raises because the call to ci._internal_get_values returns a DatetimeIndex. Is raising the desired behavior, or maybe we want to match the series values behavior?

[WillAyd]

get_values in objToJSON calls ser._internal_get_values() and gets a ndarray[dt64_naive], while cser._internal_get_values() gets a ndarray[object] of Timestamp objects. Is going through object-dtype intentional, or would it be OK to match the non-categorical behavior?

Shouldn't need an object coercion. Off the top of my head there were some bugs with datetimelike values where coercion to object would have yielded different values in some combination of the index / values, but those should be squashed now (or if not, would want to identify and correct)

[jbrockmendel]

OK. Cleanest option I've identified is to use obj.values.__array__() instead of obj._internal_get_values(). The only behavior difference is that CategoricalIndex[dt64tz] (or Series with that for its index) will no longer raise. Is that acceptable?