fillna for np.float64(np.nan)

[randomseed42]

I like pd.Float64Dtype() and other pandas datatypes, but one small glitch found in below minimal case:

import pandas as pd

df = pd.DataFrame([(2, 1), (0, 0), (0, None)], columns=('a', 'b'))
df = df.astype(pd.Float64Dtype())

print(df['b'] / df['a'])
# 0     0.5
# 1     NaN
# 2    <NA>
# dtype: Float64

print((df['b'] / df['a']).fillna(0))
# 0    0.5
# 1    NaN
# 2    0.0
# dtype: Float64

print((df['b'] / df['a']).loc[1])
# np.float64(nan)

Is there any plan to make fillna adapt to np.float64(np.nan)? Or make sure 0 / 0 in pd.Float64Dtype() can return pd.NA rather than np.float64(np.nan)?

[angela-tarantula]

I can't reproduce this on 3.0.0.dev0+2532.ga609b57b58.

[randomseed42]

I can't reproduce this on 3.0.0.dev0+2532.ga609b57b58.

Sorry I've forgot to mention the versions: Python 3.13.9, Pandas 2.3.3. Let me try in pandas v3.0.0 dev version, probably this is already fixed in upcoming version. Thanks!

[randomseed42]

I can't reproduce this on 3.0.0.dev0+2532.ga609b57b58.

Sorry I've forgot to mention the versions: Python 3.13.9, Pandas 2.3.3. Let me try in pandas v3.0.0 dev version, probably this is already fixed in upcoming version. Thanks!

import pandas as pd
# + meson compile
# INFO: autodetecting backend as ninja
# INFO: calculating backend command to run: c:\d\codespace\temp\pandas_fillna\.venv\Scripts\ninja.EXE
# [1/1] Generating write_version_file with a custom command

pd.__version__
# '3.0.0.dev0+2582.gd597079557'

df = pd.DataFrame([(2, 1), (0, 0), (0, None)], columns=('a', 'b'))
df = df.astype(pd.Float64Dtype())

print(df['b'] / df['a'])
# 0     0.5
# 1    <NA>
# 2    <NA>
# dtype: Float64

print((df['b'] / df['a']).fillna(0))
# 0    0.5
# 1    0.0
# 2    0.0
# dtype: Float64

Hooray!