When you have multiple columns to groupby and your fingers are tempted to use all column names at once in groupby syntax, don't do that. Do this instead.

Sort_in_group_multiple.py

# Define a function in which the group will be passed
def get_nlargest(group):
    
    # That group's group will be created (not complex)
    group = group.groupby('key2').value_col.sum()
    return group.nlargest(5)
  
# X Instead of doing this ↓
df.groupby(['key1', 'key2').apply(lamda group: group...)

# Do this ↓
df.groupby('key1').apply(get_nlargest)
            
'''Just pass one key and the again group inside of the function'''

Got from the pandas book, 5. Putting together, 4th book