df.fillna inplace bug

[vmirnv]

This code:

a = df()
a['f'] = [None, None]
a['s'] = [None, None]
a['t'] = [None, None]
a.loc[:,['f','s']] = a.loc[:,['f','s']].fillna(-1) #works
a.loc[:,'f'].fillna(-1, inplace=True) #also works
a.loc[:,['f','s']].fillna(-1, inplace=True) #doesn't

It's not the biggest error, but still.

[shoyer]

please showed actual and expected output

[vmirnv]

Should be like this:

    f   s   t
0   -1  -1  None
1   -1  -1  None

But instead after a.loc[:,['f','s']].fillna(-1, inplace=True) you get this:

    f   s   t
0   None    None    None
1   None    None    None

[jreback]

You are chained indexing, and operating on a copy (numpy is creating a copy here, for object type is often does this). See docs here

You should not use a chained inplace op, ever. It is pretty hard for pandas to detect this though.

Further, in general you should NOT be using object dtypes to represent numbers (which is what you get with None). NaN is the missing value marker.

In [20]: id(a.values.base)
Out[20]: 4430108432

# so this is a view
In [19]: id(a.loc[:,'f'].values.base)
Out[19]: 4430108432

# this is a copy
In [18]: id(a.loc[:,['f','s']].values.base)
Out[18]: 4431376592