BUG: Replacing NaN with None in Pandas 1.3 does not work

[pvieito]

• I have checked that this issue has not already been reported.
• I have confirmed this bug exists on the latest version of pandas.
• (optional) I have confirmed this bug exists on the master branch of pandas.

Code Sample, a copy-pastable example

Pandas 1.2

>>> import pandas as pd
>>> import numpy as np
>>> df = pd.DataFrame([0.5, np.nan])
>>> df.where(pd.notnull(df), None)
     0
0  0.5
1  None

Pandas 1.3

>>> import pandas as pd
>>> import numpy as np
>>> df = pd.DataFrame([0.5, np.nan])
>>> df.where(pd.notnull(df), None)
     0
0  0.5
1  NaN

Problem description

Replacing NaN values with None (or any other Python object) should work as in previous Pandas versions.

Expected Output

>>> import pandas as pd
>>> import numpy as np
>>> df = pd.DataFrame([0.5, np.nan])
>>> df.where(pd.notnull(df), None)
     0
0  0.5
1  None

Output of pd.show_versions()

INSTALLED VERSIONS

commit : f00ed8f
python : 3.8.10.final.0
python-bits : 64
OS : Darwin
OS-release : 21.0.0
Version : Darwin Kernel Version 21.0.0: Sun Jun 20 18:43:49 PDT 2021; root:xnu-8011.0.0.121.4~2/RELEASE_X86_64
machine : x86_64
processor : i386
byteorder : little
LC_ALL : None
LANG : es_ES.UTF-8
LOCALE : es_ES.UTF-8

pandas : 1.3.0
numpy : 1.18.5
pytz : 2021.1
dateutil : 2.8.1
pip : 21.1.3
setuptools : 57.1.0
Cython : None
pytest : None
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : 1.3.7
lxml.etree : 4.6.2
html5lib : None
pymysql : None
psycopg2 : 2.9.1 (dt dec pq3 ext lo64)
jinja2 : 2.11.3
IPython : 7.20.0
pandas_datareader: None
bs4 : 4.9.3
bottleneck : None
fsspec : None
fastparquet : None
gcsfs : None
matplotlib : 3.3.4
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : 3.0.0
pyxlsb : None
s3fs : None
scipy : 1.3.3
sqlalchemy : 1.4.20
tables : None
tabulate : None
xarray : None
xlrd : 2.0.1
xlwt : None
numba : None

[lerome]

That was apparently changed on purpose: #39761 but it's indeed not very convenient, for example we often want to replace NaN by None just before converting the dataframe back to a dict (df.to_dict) or other data structure. I'm pretty sure that this sole change breaks many existing code bases.

[rhshadrach]

One can still convert to None, however you need to convert to object dtype first:

df = pd.DataFrame([0.5, np.nan])
df = df.astype(object)
print(df.where(pd.notnull(df), None))

gives

      0
0   0.5
1  None

This behavior similar to the result of assigning using .loc/.iloc:

df = pd.DataFrame([0.5, np.nan])
df.iloc[0, 0] = None
print(df)

gives (on both 1.3.x and 1.2.x):

    0
0 NaN
1 NaN

I can understand this may be somewhat surprising, but is certainly more consistent. It also allows assignment without conversion to object, and conversion to object can involve significant performance degredation.

cc @jbrockmendel for any other thoughts.

[shahyash10]

When you're workin with json, json doesn't recognize NaNs. Conversion to object dtype would be mandatory in this scenario.
Can you please suggest some other workaround? @rhshadrach

[rhshadrach]

@yashs97 It would be helpful to understand how the workaround I posted is not sufficient.

[shahyash10]

i was asking for a workaround without converting to object dtype since like you said it involves performance degradation

[jbrockmendel]

One can still convert to None, however you need to convert to object dtype first

This is the right way to do this.

i was asking for a workaround without converting to object dtype since like you said it involves performance degradation

The 1.2 behavior referenced by the OP also converts to object, so if you want None in your Series, there is no way to do this.

[rhshadrach]

It seems to me this issue can be closed. If anyone (cc @pvieito, @lerome) would like to discuss further, please comment here and can reopen.