Lazysort

Adds a method to iterators that returns a sorted iterator over the data. The sorting is achieved lazily using a quicksort algorithm.

Available via crates.io.

Usage

extern crate lazysort;

use lazysort::Sorted;

use lazysort::SortedBy;

use lazysort::SortedPartial;

The Sorted trait adds a method sorted to all Iterator<T: Ord> which returns an iterator over the same data in default order.

The SortedBy trait adds a method sorted_by to all Iterator<T> which returns an iterator over the same data ordered according to the provided closure/function of type Fn(&T, &T) -> Ordering

The SortedPartial trait adds two methods sorted_partial_first and sorted_partial_last to all Iterator<T: PartialOrd> which returns an iterator over the same data in the default order. The difference between the two is whether non-comparable values go first or last in the results.

For example:

let data: Vec<uint> = vec![9, 1, 3, 4, 4, 2, 4];
for x in data.iter().sorted() {
	println!("{}", x);
}

Will print: 1, 2, 3, 4, 4, 4, 9

A more complex example. Sort strings by length, then in default string order:

let before: Vec<&str> = vec!["a", "cat", "sat", "on", "the", "mat"];
before.iter().sorted_by(|a, b| {
    match a.len().cmp(&b.len()) {
        Equal => a.cmp(b),
        x => x
    }
})

This returns an iterator which yields: a, on, cat, mat, sat, the.

Implementation details and performance

The algorithm is essentially the same as described in my blog post using a lazy sort as an example of Clojure's lazy sequences. But made to fit in with Rust's iterators.

The full sequence from the parent iterator is read, then each call to next returns the next value in the sorted sequence. The sort is done element-by-element so the full order is only realised by iterating all the way through to the end.

The algorithm is the quicksort, but depth-first; upon each call to next it does the work necessary to find the next item then pauses the state until the next call to next.

To test performance we compare it against sorting the full vector, using the sort function from the standard library, and also against std::collections::BinaryHeap.

First we compare what happens when sorting the entire vector:

test benches::c_heap_bench     ... bench:   3,703,166 ns/iter (+/- 454,189)
test benches::c_lazy_bench     ... bench:   3,961,047 ns/iter (+/- 603,083)
test benches::c_standard_bench ... bench:   3,093,873 ns/iter (+/- 430,401)

There are differences between the three, and not surprisingly the built-in sort is fastest.

These benchmarks are for sorting 50,000 random uints in the range 0 <= x < 1000000. Run cargo bench to run them.

So what's the point of lazy sorting? As per the linked blog post, they're useful when you do not need or intend to need every value; for example you may only need the first 1,000 ordered values from a larger set.

Comparing the lazy approach data.iter().sorted().take(x) vs a standard approach of sorting a vector then taking the first x values gives the following.

The first 1,000 out of 50,000:

test benches::a_heap_bench     ... bench:     366,767 ns/iter (+/- 55,393)
test benches::a_lazy_bench     ... bench:     171,923 ns/iter (+/- 52,784)
test benches::a_standard_bench ... bench:   3,055,734 ns/iter (+/- 348,407)

The lazy approach is quite a bit faster; this is due to the 50,000 only being sorted enough to identify the first 1,000, the rest remain unsorted. BinaryHeap is also quite fast, for the same reason.

The first 10,000 out of 50,000:

test benches::b_heap_bench     ... bench:   1,126,774 ns/iter (+/- 156,833)
test benches::b_lazy_bench     ... bench:     993,954 ns/iter (+/- 208,188)
test benches::b_standard_bench ... bench:   3,054,598 ns/iter (+/- 285,970)

The lazy approach is still faster in this situation.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.