first, I know the strings in column b can be used directly for sorting. My function aFucntion seems to be so stupid, because it is just a simplified one to present the problem. In my real situation, it actually produces a list of some English Characters, which then be used to sort words in my native language in a customed way.

Then, I know LIST is not hashable

# coding = utf-8
import pandas as pd

def aFucntion(var):
    return list(var)

a = {'a': [1, 2],
     'b': ['Python', 'Java']
     }

df =pd.DataFrame(a)

print(df)
print('\n')


df['fun'] = df['b'].map(aFucntion)


df1 = df.sort_values(['b']) # this works
print(df1)
print('\n')

df2 = df.sort_values(['fun'])  # this works
print(df2)
print('\n')

df3 = df.sort_values(['b', 'fun'])  # this yields `TypeError: unhashable type: 'list'`
# but why df.sort_values(['fun'])  works?

print(df3)
print('\n')

df['fun'] = df['b'].map(lambda e: tuple(aFucntion(e)))
df4 = df.sort_values(['b', 'fun'])  # this works
print(df4)
print('\n')

Comment From: TomAugspurger

You can see here that sort_values treats the case where length(by) is greater than one differently than when it's equal to one. When you pass a list of columns to sort by, it has to go through our factorization routines, which requires hashable values.

Comment From: michaelwebb7

@TomAugspurger I just came up against this bug – being able to sort by two columns seems a very basic feature that ought to work. Is it possible to fix this? I created a temporary workaround for myself by making a new column that stringified and concatenated the 2 columns I cared about, but it would be great if this just worked out of the box.

Comment From: TomAugspurger

@michaelwebb7 you're able to sort by multiple columns. You aren't able to sort by something that isn't hashable.

Comment From: TomAugspurger

sort_values needs the values of a column to be hashable when sorting by multiple columns. The original example could be fixed by converting the column of (unhashable) lists to a column of tuples.

In [10]: df['fun2'] = df['fun'].map(tuple)

In [11]: df
Out[11]:
   a       b                 fun                fun2
0  1  Python  [P, y, t, h, o, n]  (P, y, t, h, o, n)
1  2    Java        [J, a, v, a]        (J, a, v, a)

In [12]: df.sort_values(['a', 'fun2'])
Out[12]:
   a       b                 fun                fun2
0  1  Python  [P, y, t, h, o, n]  (P, y, t, h, o, n)
1  2    Java        [J, a, v, a]        (J, a, v, a)

Comment From: m-martin-j

A workaround that helped me is

df2 = df.applymap(tuple) df2.sort_values(['b', 'fun'])

to sort by more than one column