Question

I am looking for a function ideally in pure python that is similar to numpy.argsort in that it returns only a list of sorted indices while leaving the original arrays untouched, yet it needs to be able to sort on data contained in multiple arrays.

Example:

>>> names = ['xavier', 'bob', 'billy', 'jene', 'samson']
>>> ages = [15, 32, 63, 32, 15]
>>>indexes = sort by ages and then by names
[4, 0, 1, 3, 2]
>>> for i in indexes:
>>>    print "Name", names[i]
>>>    print "Age", ages[i]

The sorting function cannot create extra data structures, meaning list comprehension or functions like zip are out of the question. Each array consists of 5 million objects, generating zipped version of the arrays explodes the memory requirements by a factor of at least 3. Using list comprehension such as sorted(..key=lambda x:(names[x],ages[x])) causes a slowdown such as the sort takes over a minute to complete (and the memory requirements to create these intermediary tuples)

So far, as long as I only want to sort on a single array it is fast enough, however since the indices list does not know about the other arrays, I am unable to call multiple “sort” operations as I would if I had a zipped version of the two lists.

Answer 1

I created my own solution that works great.

Given the following data set:

groups = reversed(range(5000000))
ages = [random.randrange(0, 120) for x in groups]
names = ['foobar-%d' % random.randrange(0, 5000) for x in groups]

columns = dict(names=names,ages=ages,groups=groups)

def sort_on(col):
    idxs = range(len(columns[col]))
    idxs.sort(key=lambda x:columns[col][x])
    return idxs