import numpy as np
import pandas as pd
ii = pd.interval_range(1, 10)
ser = pd.Series(ii).copy()
mask = np.zeros(ser.shape, dtype=bool)
mask[1] = True
ii.insert(1, np.nan) # <- IntervalDtype[float64]
ii.putmask(mask, np.nan) # <- IntervalDtype[float64]
ser.where(mask, np.nan) # <- ValueError
ser[1] = np.nan <- coerces to object
ATM dtypes.cast.ensure_dtype_can_hold_na is incorrect for EA dtypes that cannot hold NA values (the IntervalDtype[int] example is the only one that comes to mind).
We have special-casing for IntervalDtype in Index._find_common_type_compat which is why the Index methods above behave well. Short-term, we can move that special-casing into ensure_dtype_can_hold_na. Longer-term, we need a way for EADtype subclasses to specify this behavior.
This could be seen as a special case of xref #24246
cc @jorisvandenbossche @TomAugspurger
Comment From: jorisvandenbossche
There is already a _can_hold_na attribute on the ExtensionDtype. It's of course not fully the same (it doesn't distinguish between different types of NAs, just whether it can hold any type of NA), but for the specific use case of IntervalDtype[int] not being able to store NAs that could be sufficient?
Comment From: jbrockmendel
There is already a _can_hold_na attribute on the ExtensionDtype
The question isn't "can this store an NA", it's "given that it can't, what's the closest dtype that can?"
Comment From: jbrockmendel
Just discovered there are a couple places in core.indexing where we call maybe_promote (which is annotated as only taking numpy dtypes) with potentially-EA dtypes. Inside maybe_promote we have special handling for CategoricalDtype (which is reached in the tests). And outside maybe_promote in the indexing code with have curr_dtype = getattr(self.obj.dtype, "numpy_dtype", self.obj.dtype) which looks like a workaround to support MaskedDtypes here.
Comment From: jorisvandenbossche
Longer-term, we need a way for EADtype subclasses to specify this behavior.
Or alternatively: longer term, every EA can hold NAs, or otherwise raise an error if we try to put NAs in it (this also relates to the proposal to preserve dtypes in setitem-like operations).
(I know this doesn't solve the specific issue for Interval[int], but if that's the currently the only one that gives a wrong result in ensure_dtype_can_hold_na, that might not be worth making a generic property for it on the EA class/dtype)