Add key to sorting functions

[hayd]

Many python functions (sorting, max/min) accept a key argument, perhaps they could in pandas too.

.

The terrible motivating example was this awful hack from this question.... for which maybe one could do

df.sort_index(key=lambda t: literal_eval(t[1:-1]))

This would still be an awful awful hack, but a slightly less awful one.

[SethMMorton]

Here's a specific use case that came up on StackOverflow recently.

from pandas import DataFrame
df = DataFrame({'a':range(5)},index=['0hr', '128hr', '72hr', '48hr', '96hr'])
print(df.sort())

This returns

       a
0hr    0
128hr  1
48hr   3
72hr   2
96hr   4

The user would have liked to sort the index "naturally" using natsort, i.e.:

from natsort import natsort_keygen
print(df.sort(key=natsort_geygen()))

which would have returned

       a
0hr    0
48hr   3
72hr   2
96hr   4
128hr  1

[jreback]

Here's the way to do this (further you could use a CateogricalIndex to have custom sorting).

In [67]: df = DataFrame({'a':range(5)},index=pd.to_timedelta(['0h', '128h', '72h', '48h', '96h']))

In [68]: df
Out[68]: 
                 a
0 days 00:00:00  0
5 days 08:00:00  1
3 days 00:00:00  2
2 days 00:00:00  3
4 days 00:00:00  4

In [69]: df.sort_index()
Out[69]: 
                 a
0 days 00:00:00  0
2 days 00:00:00  3
3 days 00:00:00  2
4 days 00:00:00  4
5 days 08:00:00  1

[SethMMorton]

Agreed, in this specific example they could use timedeltas. But what if the index labels were something like "version1", "version2" ... "version10", etc. or something else that does not directly correspond with a builtin, sortable datatype? I think I'm looking for a more general solution.

I'm not really sure how to use CategoricalIndex to do custom sorting... I'm getting a bit lost in the docs. Is that considered by the pandas community to be the correct way to do custom sorting?

[jreback]

see #9741 (big, long), but short answer is yes.

[SethMMorton]

If I understand properly, in order for Categorical to work properly the sorting order for all elements must be defined at the time of creation of the DataFrame. What if additional elements are added that were not present in the initial definition?

Based on your suggestions, I am getting the impression that key will not be added to sort. Am I reading between the lines correctly?