Python Pandas Dataframe idxmax is so slow. Alternatives?

Question

I'm trying to select rows out of groups by max value using df.loc[df.groupby(keys)['column'].idxmax()].

I'm finding, however, that df.groupby(keys)['column'].idxmax() takes a really long time on my dataset of about 27M rows. Interestingly, running df.groupby(keys)['column'].max() on my dataset takes only 13 seconds while running df.groupby(keys)['column'].idxmax() takes 55 minutes. I don't understand why returning the indexes of the rows takes 250 times longer than returning a value from the row. Maybe there is something I can do to speed up idxmax?

If not, is there an alternative way of selecting rows out of groups by max value that might be faster than using idxmax?

For additional info, I'm using two keys and sorted the dataframe on those keys prior to the groupby and idxmax operations. Here's what it looks like in Jupyter Notebook:

import pandas as pd

df = pd.read_csv('/data/Broadband Data/fbd_us_without_satellite_jun2019_v1.csv', encoding='ANSI', \
    usecols=['BlockCode', 'HocoNum', 'HocoFinal', 'TechCode', 'Consumer', 'MaxAdDown', 'MaxAdUp'])

%%time
df = df[df.Consumer == 1]
df.sort_values(['BlockCode', 'HocoNum'], inplace=True)
print(df)

              HocoNum                  HocoFinal        BlockCode  TechCode
    4631064    130077                  AT&T Inc.   10010201001000        10   
    4679561    130077                  AT&T Inc.   10010201001000        11   
    28163032   130235     Charter Communications   10010201001000        43   
    11134756   131480  WideOpenWest Finance, LLC   10010201001000        42   
    11174634   131480  WideOpenWest Finance, LLC   10010201001000        50   
    ...           ...                        ...              ...       ...   
    15389917   190062          Broadband VI, LLC  780309900000014        70   
    10930322   130081    ATN International, Inc.  780309900000015        70   
    15389918   190062          Broadband VI, LLC  780309900000015        70   
    10930323   130081    ATN International, Inc.  780309900000016        70   
    15389919   190062          Broadband VI, LLC  780309900000016        70   

              Consumer  MaxAdDown  MaxAdUp  
    4631064          1        6.0    0.512  
    4679561          1       18.0    0.768  
    28163032         1      940.0   35.000  
    11134756         1     1000.0   50.000  
    11174634         1     1000.0   50.000  
    ...            ...        ...      ...  
    15389917         1       25.0    5.000  
    10930322         1       25.0    5.000  
    15389918         1       25.0    5.000  
    10930323         1       25.0    5.000  
    15389919         1       25.0    5.000  

    [26991941 rows x 7 columns]
    Wall time: 21.6 s

%time df.groupby(['BlockCode', 'HocoNum'])['MaxAdDown'].max()

    Wall time: 13 s
    BlockCode        HocoNum
    10010201001000   130077       18.0
                     130235      940.0
                     131480     1000.0
    10010201001001   130235      940.0
    10010201001002   130077        6.0
                                 ...  
    780309900000014  190062       25.0
    780309900000015  130081       25.0
                     190062       25.0
    780309900000016  130081       25.0
                     190062       25.0
    Name: MaxAdDown, Length: 20613795, dtype: float64

%time df.groupby(['BlockCode', 'HocoNum'])['MaxAdDown'].idxmax()

Wall time: 55min 24s
BlockCode        HocoNum
10010201001000   130077      4679561
                 130235     28163032
                 131480     11134756
10010201001001   130235     28163033
10010201001002   130077      4637222
                              ...   
780309900000014  190062     15389917
780309900000015  130081     10930322
                 190062     15389918
780309900000016  130081     10930323
                 190062     15389919
Name: MaxAdDown, Length: 20613795, dtype: int64

You'll see in the very first rows of data there are two entries for AT&T in the same BlockCode, one for MaxAdDown of 6Mbps and one for 18Mbps. I want to keep the 18Mbps row and drop the 6Mbps row, so that there is one row per company per BlockCode that has the the maximum MaxAdDown value. I need the entire row, not just the MaxAdDown value.

Answer 1

sort and drop duplicates:

df.sort('MaxAdDown').drop_duplicates(['BlockCode', 'HocoNum'], keep='last')