Description
dti = pd.DatetimeIndex([pd.NaT, "2016-01-01"], tz="UTC")
tdi = dti - dti[1]
pi = dti.tz_localize(None).to_period("D")
ci = pi.astype("category")
ii = pd.IntervalIndex([None])
pd.Series(dti).fillna("foo") # <- casts
pd.Series(tdi).fillna("foo") # <- raises
pd.Series(pi).fillna("foo") # <- casts, but untested
pd.Series(ci).fillna("foo") # <- raises
pd.Series(ii).fillna("foo") # <- raises, untested
ATM fillna behavior with incompatible values is pretty inconsistent. With numpy dtypes (except dt64 and td64) we always cast. With dt64 and dt64tz we always cast.
With td64 we always raise, but we only have tests for the value being an integer. Back circa 2018 we used to coerce integers to timedelta, so pd.Series(tdi).fillna(1) would interpret the 1 as pd.Timedelta(seconds=1) (not nanos!). We deprecated that coercion and now raise, but I don't think there was much thought given to whether to raise versus cast to object.
With Categorical we raise. We have 2 tests specific to that.
With PeriodDtype we cast and with IntervalDtype we raise, but we have tests for neither.
I don't have a particularly strong opinion on this, but would like to be consistent.