Add DataFrame.sort and Column.sort

[MarcoGorelli]

Feedback I received when showing this to people is that

col.get_rows(col.sorted_indices())

isn't ergonomic

It's also a performance footgun

In [1]: df = pl.DataFrame({'a': np.random.randint(0, 4, size=100_000_000)})

In [2]: %timeit df[df['a'].arg_sort()]
2.56 s ± 80.3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [3]: %timeit df.sort('a')
700 ms ± 21.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

[kkraus14]

That performance difference for Polars is quite surprising. Does df.__getitem__ do a lot of extra bounds checking or something else that could be bypassed in the case where you know the indices being passed to it are generated with guarantees?

[rgommers]

This sounds very reasonable to add, LGTM.

It would probably be useful to edit the sorted_indices docstrings, since it contains:

        If you need to sort the DataFrame, you can simply do::

            df.get_rows(df.sorted_indices(keys))

instead, that can now refer to the sort method

[MarcoGorelli]

That performance difference for Polars is quite surprising. Does df.__getitem__ do a lot of extra bounds checking or something else that could be bypassed in the case where you know the indices being passed to it are generated with guarantees?

there are no guarantees in this case

anyway, you can observe the same in numpy:

In [14]: arr = np.random.randint(0, 4, size=100_000_000)

In [15]: %time arr[np.argsort(arr)]
CPU times: user 767 ms, sys: 2.13 s, total: 2.89 s
Wall time: 2.9 s
Out[15]: array([0, 0, 0, ..., 3, 3, 3])

In [16]: %time np.sort(arr)
CPU times: user 253 ms, sys: 187 ms, total: 440 ms
Wall time: 439 ms
Out[16]: array([0, 0, 0, ..., 3, 3, 3])