API/BUG: pd.array(float_ndarray_with_nans) casts np.nan to pd.NA

[jbrockmendel]

I was under the impression we meant to keep these as distinct concepts

[jbrockmendel]

setting data[-3] = np.nan also gets cast to pd.NA

[phofl]

Docs sound like this is intended? https://pandas.pydata.org/docs/reference/api/pandas.array.html

Changed in version 1.2.0: Pandas now also infers nullable-floating dtype for float-like input data

[jbrockmendel]

@jorisvandenbossche is the one to ask on this

[jorisvandenbossche]

Yes, this is not a bug but intentional behaviour.

See the 4th bullet points in the top post of the PR implementing it (#34307), but this just mentions it. Some actual discussion about is scattered through the issue on this: #32265

The main reason is that any code that currently deals with numpy arrays and pandas (also internally in pandas) assumes that NaNs are treated as missing data. So for compatibility with this (preserving the "missing" semantics), we convert any NaN to NA by default when converting numpy array to a masked/nullable array.

Now, this certainly is a topic that still needs more attention (eg adding an option to be able to preserve NaNs on input instead of converting to NA, we still need to discuss how to treat NaNs that get introduced by computations, default conversion to numpy, ..)

[jbrockmendel]

Thanks @jorisvandenbossche. Is there (supposed to be?) a way to actually set data[3] = np.nan and actually get np.nan instead of pd.NA?

[simonjayhawkins]

duplicate of/covered by #32265

[mroeschke]

Does look similar to #32265, continued discussion can happen on that issue