pd.NA in object dtype

[simonjayhawkins]

extract from #32075 (comment)

If we want to handle pd.NA in object dtype better, we will need to start using masks as well, and not rely on numpy behaviour.

For example, also this is wrong:

In [9]: pd.Series([1, pd.NA], dtype=object) >= 1
Out[9]: 
0     True
1    False
dtype: bool

Related issues:

• Series.replace fails to replace value #32075 (replace not working)
• BUG: truncated repr with pd.NA in object dtype column shows "NaN" #33065 (wrong truncated repr)
• BUG: pd.NA acts differently when inside/outside a series/dataframe with object dtype #33066 (wrong comparison operations)
• BUG: Cannot index into DataFrame with Nullable Integer as index  #34497 (indexing with duplicates)

[ianbtr]

It looks like pd.NA is dropped entirely when concatenating two dataframes with object dtype, and this is very similar to #33065:

In[4]: df1 = pd.DataFrame(np.full((1, 1), pd.NA))
In[5]: df2 = pd.DataFrame(df1)
In[6]: df1
Out[6]: 
      0
0  <NA>
In[7]: df2
Out[7]: 
      0
0  <NA>
In[8]: pd.concat([df1, df2])
Out[8]: 
     0
0  NaN
0  NaN

This seems to be due to _get_empty_dtype_and_na and JoinUnit.get_reindexed_values in concat.py. The former selects np.nan for anything of dtype 'object'. In this case (and probably many others), 'object' implicitly has a null-value of np.nan.

If the ObjectBlock contained a mask for pd.NA, then JoinUnit.get_reindexed_values could apply it as necessary, and several other functions could use it as well.

Would this work?

[eddy-geek]

Not quite related -- I am surprised that

• mix of int and pd.NA defaults to object dtype
• convert_dtypes leaves it as object

Are these working as designed, already being ironed out, or should I open issues?

df = pd.DataFrame([7,8,9,pd.NA])
print(df)
print('auto        ', df.dtypes)
print('auto convert', df.convert_dtypes().dtypes)

dfi = pd.DataFrame([7,8,9,pd.NA], dtype='Int64')
print(dfi)
print('Int64        ', dfi.dtypes)
print('Int64 convert', dfi.convert_dtypes().dtypes)

print('pandas', pd.__version__)

gives:

      0
0     7
1     8
2     9
3  <NA>
auto         0    object
dtype: object
auto convert 0    object
dtype: object

      0
0     7
1     8
2     9
3  <NA>
Int64         0    Int64
dtype: object
Int64 convert 0    Int64
dtype: object
pandas 1.0.3

[simonjayhawkins]

• mix of int and pd.NA defaults to object dtype

since pd.NA is experimental, changing the constructor to default to the best possible dtypes using dtypes supporting pd.NA seems reasonable.

• convert_dtypes leaves it as object

from https://pandas.pydata.org/docs/dev/reference/api/pandas.DataFrame.convert_dtypes.html

Convert columns to best possible dtypes using dtypes supporting pd.NA.

so again these seems to be a reasonable expectation.

however, from https://pandas.pydata.org/docs/dev/user_guide/missing_data.html?highlight=convert_dtypes#conversion

If you have a DataFrame or Series using traditional types that have missing data
represented using np.nan, there are convenience methods
:meth:~Series.convert_dtypes in Series and :meth:~DataFrame.convert_dtypes
in DataFrame that can convert data to use the newer dtypes for integers, strings and
booleans

so it maybe that the convert_dtypes docstring should also be more explicit about the conversion applies to np.nan

or should I open issues?

These two issues could be discussed/addressed independently, so if you could report these as two independent issues, that'll be great.