#priority-queue #queue #priority

nightly priq

Array implementation of the min/max heap

4 releases

0.2.0 Feb 14, 2022
0.1.6 Feb 12, 2022
0.1.5 Feb 10, 2022
0.1.1 Feb 10, 2022
0.1.0 Feb 10, 2022

#1069 in Data structures

48 downloads per month
Used in clocks

MIT license

45KB
496 lines

priq

Priority queue (min/max heap) using raw binary heap.

PriorityQueue is built using raw array for efficient performance.

There are two major reasons what makes this PriorityQueue different from other binary heap implementations currently available:

1 - Allows data ordering to scores with PartialOrd. - Every other min-max heap requires total ordering of scores (e.g. should implement Ord trait). This can be an issue, for example, when you want to order items based on a float scores, which doesn't implement Ord trait. - Because of partial ordering, non-comparable values are thrown in the end of the queue. One will see non-comparable values only after all the comparable elements have been pop-ed. - You can read about Rust's implementation or Ord, PartialOrd and what's the different here

2 - Separation of score and item you wish to store. - This frees enforcement for associated items to implement any ordering. - Makes easier to evaluate items' order.

3 - Equal scoring items are stored at first available free space. - This gives performance boost for large number of entries.

4 - Easy to use!

You can read more about this crate on my blog

Implementation

A Min-Max Heap with designated arguments for score and associated item!

A Default implementation is a Min-Heap where the top node (root) is the lowest scoring element:

                    10
                   /  \
                58      70
               /  \    /  \
             80   92  97   99

The value of Parent Node is small than Child Node.

Every parent node, including the top (root) node, is less than or equal to equal to the right child.

PriorityQueue allows duplicate score/item values. When you putthe item with a similar score that’s already in the queue new entry will be stored at the first empty location in memory. This gives an incremental performance boost (instead of resolving by using the associated item as a secondary tool to priority evaluation). Also, this form of implementation doesn’t enforce for the element T to have any implemented ordering. This guarantees that the top node will always be of minimum value.

You can initialize an empty PriorityQueue and later add items:

use priq::PriorityQueue;

// create queue with `usize` key and `String` elements
let pq: PriorityQueue<usize, String> = PriorityQueue::new();

Or you can heapify a Vec and/or a slice:

use priq::PriorityQueue;

let pq_from_vec = PriorityQueue::from(vec![(5, 55), (1, 11), (4, 44)]);
let pq_from_slice = PriorityQueue::from([(5, 55), (1, 11), (4, 44)]);

Partial Ordering

Because priq allows score arguments that only implement PartialOrd, elements that can't be compared are evaluated and are put in the back of the queue:

use priq::PriorityQueue;

let mut pq: PriorityQueue<f32, isize> = PriorityQueue::new();

pq.put(1.1, 10);
pq.put(f32::NAN, -1);
pq.put(2.2, 20);
pq.put(3.3, 30);
pq.put(f32::NAN, -3);
pq.put(4.4, 40);

(1..=4).for_each(|i| assert_eq!(i * 10, pq.pop().unwrap().1));

// NAN scores will not have deterministic order
// they are just stored after all the comparable scores
assert!(0 > pq.pop().unwrap().1);
assert!(0 > pq.pop().unwrap().1);

Time

The standard usage of this data structure is to put an element to the queue and pop to remove the top element and peek to check what’s the top element in the queue. The stored structure of the elements is a balanced tree realized using an array with a contiguous memory location. This allows maintaining a proper parent-child relationship between put-ed items.

Runtime complexity with Big-O Notation:

method Time Complexity
put O(log(n))
pop O(log(n))
peek O(1)

You can also iterate over elements using for loop but the returned slice will not be properly order as the heap is re-balanced after each insertion and deletion. If you want to grab items in a proper priority call pop in a loop until it returns None.

Custom struct

What if you want to custom struct without having a separate and specific score? You can pass the struct’s clone as a score and as an associated value, but if in this kind of scenario I’d recommend using BinaryHeap as it better fits the purpose.

Min-Heap

If instead of Min-Heap you want to have Max-Heap, where the highest-scoring element is on top you can pass score using Reverse or a custom [Ord] implementation can be used to have custom prioritization logic.

Example

use priq::PriorityQueue;
use std::cmp::Reverse;

let mut pq: PriorityQueue<Reverse<u8>, String> = PriorityQueue::new();

pq.put(Reverse(26), "Z".to_string());
pq.put(Reverse(1), "A".to_string());

assert_eq!(pq.pop().unwrap().1, "Z");

Merging and Combining

You can merge another priority queue to this one. Right hand side priority queue will be drained into the left hand side priority queue.

Examples

use priq::PriorityQueue;

let mut pq1 = PriorityQueue::from([(5, 55), (6, 66), (3, 33), (2, 22)]);
let mut pq2 = PriorityQueue::from([(4, 44), (1, 11)]);

pq1.merge(&mut pq2);
// at this point `pq2` is empty

assert_eq!(6, pq1.len());
assert_eq!(11, pq1.peek().unwrap().1);

You can also use + operator to combine two priority queues. Operands will be intact. New priority queue will be build from cloning and merging them.

Example

use priq::PriorityQueue;

let pq1 = PriorityQueue::from([(5, 55), (1, 11), (4, 44), (2, 22)]);
let pq2 = PriorityQueue::from([(8, 44), (1, 22)]);

let res = pq1 + pq2;

assert_eq!(6, res.len());
assert_eq!(11, res.peek().unwrap().1);

Performance

This are the benchmark results for priq::PriorityQueue:

priq benches median nanosecs std.dev
pq_pop_100 146 ns/iter (+/- 1)
pq_pop_100k 291,818 ns/iter (+/- 5,686)
pq_pop_10k 14,129 ns/iter (+/- 39)
pq_pop_1k 1,646 ns/iter (+/- 32)
pq_pop_1mil 16,517,047 ns/iter (+/- 569,128
pq_put_100 488 ns/iter (+/- 21)
pq_put_100k 758,422 ns/iter (+/- 13,961)
pq_put_100k_wcap 748,824 ns/iter (+/- 7,926)
pq_put_10k 80,668 ns/iter (+/- 1,324)
pq_put_1k 8,769 ns/iter (+/- 78)
pq_put_1mil 6,728,203 ns/iter (+/- 76,416)
pq_put_1mil_wcap 6,622,341 ns/iter (+/- 77,162)

How it compares to std::collections::BinaryHeap:

BinaryHeap benches median nanosecs std.dev
bh_pop_100 272 ns/iter (+/- 90)
bh_pop_100k 171,071 ns/iter (+/- 6,131)
bh_pop_10k 13,904 ns/iter (+/- 130)
bh_pop_1k 1,847 ns/iter (+/- 6)
bh_pop_1mil 8,772,066 ns/iter (+/- 611,613)
bh_push_100 857 ns/iter (+/- 50)
bh_push_100k 943,465 ns/iter (+/- 108,698)
bh_push_10k 92,807 ns/iter (+/- 7,930)
bh_push_1k 8,606 ns/iter (+/- 639)
bh_push_1mil 12,946,815 ns/iter (+/- 900,347)

Project is distributed under the MIT license. Please see the LICENSE for more information.

Dependencies

~310KB