BUG: NumpyEA[uint8].take([-1], allow_fill=True) nonsense result

[jbrockmendel]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

arr = pd.Series([1, 2, 3], dtype="uint8").array

>>> arr.take([-1], allow_fill=True)
<NumpyExtensionArray>
[nan]
Length: 1, dtype: uint8

Issue Description

The actual backing _ndarray is float64 but the result thinks it is still uint8.

Expected Behavior

N/A

Installed Versions

Replace this line with the output of pd.show_versions()

[rhshadrach]

I'm seeing the same with other sizes, e.g. int64.

[aijams]

The issue appears to be one of violated assumptions:

The take function from _mixins.py assumes that the output of its call to take from algorithms.py has the same dtype as the NDArrayBackedExtensionArray object that the _mixins take is originally called on. This is not always the case, since algorithms.take promotes integer types to floating-point types to accommodate NaN values when they occur.

Location of the bug (in _mixins.py, line 176):

        new_data = take(
            self._ndarray,
            indices,
            allow_fill=allow_fill,
            fill_value=fill_value,
            axis=axis,
        )
        return self._from_backing_data(new_data)

[aijams]

I propose setting _ndarray.dtype on self (of type NDArrayBackedExtensionArray) to the dtype of the output new_data of the sub-call to take before assigning the array data from new_data to self and returning.

This would ensure that the dtypes match between self and new_data in all cases and ultimately that the dtype reported by self is accurate.

[jbrockmendel]

@aijams i opened the issue instead of fixing it myself to give someone the chance to learn from the process of solving it. Using AI defeats the purpose.

[aijams]

Thanks for your time in giving feedback. I didn't use AI in solving this, but what makes you think I was using AI?

[jbrockmendel]

The writing style looked like it, but I'm happy to be wrong. A PR would be welcome.

[aijams]

take