Pandas “fillna” ignores the “inplace=True” keyword

Question

I have a DataFrame (fairly large, hard to reproduce &c), for which I observe this behavior:

>>> df.info(verbose=True,memory_usage=True,null_counts=True)
<class 'pandas.core.frame.DataFrame'>
Int64Index: 49841 entries, 0 to 49878
Data columns (total 70 columns):
...
channel                                25101 non-null object
...
dtypes: bool(10), datetime64[ns](6), float64(2), int64(32), object(20)
memory usage: 23.7+ MB
>>> df.channel.fillna("Unknown",inplace=True)
>>> df.info(verbose=True,memory_usage=True,null_counts=True)
<class 'pandas.core.frame.DataFrame'>
Int64Index: 49841 entries, 0 to 49878
Data columns (total 70 columns):
...
channel                                25101 non-null object
...
dtypes: bool(10), datetime64[ns](6), float64(2), int64(32), object(20)
memory usage: 23.7+ MB

IOW, it appears that df.channel.fillna("Unknown",inplace=True) has no effect.

How can that be? Is this a bug? What am I doing wrong?!

PS. Summary from the comments:

• df.is_copy is None
• df._is_view is False
• channel is a column, not an attribute, because it is listed by info

Answer 1

From the documentation :

You can use attribute access to modify an existing element of a Series or column of a DataFrame, but be careful; if you try to use attribute access to create a new column, it fails silently, creating a new attribute rather than a new column.

We suspect you assigned df.channel first, then df['channel'] and this creates the unexpected behavior.

Answer 2

The reason turned out to be the following sqlalchemy query:

select *
from table1 
join table2 
on table1.id = table2.id

The resulting DF has two columns named id and a total havoc ensues.

Solution:

select *
from table1 
join (select id as id2, ... from table2) t2 
on table1.id = t2.id2