API: Should isna return a copy for nullable types?

[mzeitlin11]

This came up in #40651, seems worth resolving. Right now BaseMaskedArray just directly returns self._mask in the implementation of isna. This could lead to some unintuitive user-facing behavior like:

>>> ser = pd.Series([1, 1, 1], dtype="Int64")
>>> mask = ser.isna()
>>> mask[:] = True
>>> ser
0    <NA>
1    <NA>
2    <NA>
dtype: Int64

Seems like isna should return a copy instead to avoid this. One potential downside is losing an API-supported way to obtain 0-cost access to the mask (xref #34873).

[jreback]

yeah this should return a copy

aside from that we could implement public methods that provide access with 0 copy

[jbrockmendel]

could do a readonly view

[mzeitlin11]

Readonly view seems like a good solution, only concern would be the inconsistency of

In [1]: import pandas as pd

In [2]: ser = pd.Series([1, 1])

In [3]: mask = ser.isna()

In [4]: mask[:] = True

In [5]: ser = pd.Series([1, 1], dtype="Int64")

In [6]: mask = ser.isna()

In [7]: mask[:] = True   # raises, ValueError: assignment destination is read-only

Are readonly views returned anywhere else? Copy-on-write view would be nice, but don't think that's easy w numpy.

[jorisvandenbossche]

Let's leave possible read-only / copy-on-write optimizations for later. The fact that it right now doesn't return a copy is simply a bug. Modifying the resulting array shouldn't modify the original.

BTW another question here is whether isna() shouldn't rather return a nullable boolean array instead of a numpy array (I know that the result of isna() will never have missing elements by definition, but still we typically return new nullable arrays from methods on nullable arrays, except for where there is an explicit conversion to numpy).