Crate im[][src]

Immutable Data Structures for Rust

This library implements several of the more commonly useful immutable data structures for Rust.

Why Would I Want This?

While immutable data structures can be a game changer for other programming languages, the most obvious benefit - avoiding the accidental mutation of data - is already handled so well by Rust's type system that it's just not something a Rust programmer needs to worry about even when using data structures that would send a conscientious Clojure programmer into a panic.

Immutable data structures offer other benefits, though, some of which are useful even in a language like Rust. The most prominent is structural sharing, which means that if two data structures are mostly copies of each other, most of the memory they take up will be shared between them. This implies that making copies of an immutable data structure is cheap: it's really only a matter of copying a pointer and increasing a reference counter, where in the case of Vec you have to allocate the same amount of memory all over again and make a copy of every element it contains. For immutable data structures, extra memory isn't allocated until you modify either the copy or the original, and then only the memory needed to record the difference.

Another goal of this library has been the idea that you shouldn't even have to think about what data structure to use in any given situation, until the point where you need to start worring about optimisation - which, in practice, often never comes. Beyond the shape of your data (ie. whether to use a list or a map), it should be fine not to think too carefully about data structures - you can just pick the one that has the right shape and it should have acceptable performance characteristics for every operation you might need. Specialised data structures will always be faster at what they've been specialised for, but im aims to provide the data structures which deliver the least chance of accidentally using them for the wrong thing.

For instance, Vec beats everything at memory usage, indexing and operations that happen at the back of the list, but is terrible at insertion and removal, and gets worse the closer to the front of the list you get. VecDeque adds a little bit of complexity in order to make operations at the front as efficient as operations at the back, but is still bad at insertion and especially concatenation. Vector adds another bit of complexity, and could never match Vec at what it's best at, but in return every operation you can throw at it can be completed in a reasonable amount of time - even normally expensive operations like copying and especially concatenation are reasonably cheap when using a Vector.

It should be noted, however, that because of its simplicity, Vec actually beats Vector even at its strongest operations at small sizes, just because modern CPUs are hyperoptimised for things like copying small chunks of contiguous memory - you actually need to go past a certain size (usually in the vicinity of one or two hundred elements) before you get to the point where Vec isn't always going to be the fastest choice. Vector attempts to overcome this by basically just being a couple of Vecs at small sizes, until it grows big enough to warrant a more traditional immutable tree structure, but even this involves a little bit of overhead.

The maps - HashMap and OrdMap - generally perform similarly to their equivalents in the standard library, but tend to run a bit slower on the basic operations (HashMap is almost neck and neck with its counterpart, while OrdMap currently tends to run 2-3x slower). On the other hand, they offer the cheap copy and structural sharing between copies that you'd expect from immutable data structures.

In conclusion, the aim of this library is to provide a safe default choice for the most common kinds of data structures, allowing you to defer careful thinking about the right data structure for the job until you need to start looking for optimisations - and you may find, especially for larger data sets, that immutable data structures are still the right choice.

Data Structures

We'll attempt to provide a comprehensive guide to the available data structures below.

Performance Notes

"Big O notation" is the standard way of talking about the time complexity of data structure operations. If you're not familiar with big O notation, here's a quick cheat sheet:

O(1) means an operation runs in constant time: it will take the same time to complete regardless of the size of the data structure.

O(n) means an operation runs in linear time: if you double the size of your data structure, the operation will take twice as long to complete; if you quadruple the size, it will take four times as long, etc.

O(log n) means an operation runs in logarithmic time: for log2, if you double the size of your data structure, the operation will take one step longer to complete; if you quadruple the size, it will need two steps more; and so on. However, the data structures in this library generally run in log64 time, meaning you have to make your data structure 64 times bigger to need one extra step, and 4096 times bigger to need two steps. This means that, while they still count as O(log n), operations on all but really large data sets will run at near enough to O(1) that you won't usually notice.

O(n log n) is the most expensive operation you'll see in this library: it means that for every one of the n elements in your data structure, you have to perform log n operations. In our case, as noted above, this is often close enough to O(n) that it's not usually as bad as it sounds, but even O(n) isn't cheap and the cost still increases logarithmically, if slowly, as the size of your data increases. O(n log n) basically means "are you sure you need to do this?"

O(1)* means 'amortised O(1),' which means that an operation usually runs in constant time but will occasionally be more expensive: for instance, Vector::push_back, if called in sequence, will be O(1) most of the time but every 64th time it will be O(log n), as it fills up its tail chunk and needs to insert it into the tree. Please note that the O(1) with the asterisk attached is not a common notation; it's just a convention I've used in these docs to save myself from having to type 'amortised' everywhere.

Lists

Lists are sequences of single elements which maintain the order in which you inserted them. The only list in this library is Vector, which offers the best all round performance characteristics: it's pretty good at everything, even if there's always another kind of list that's better at something.

Type Algorithm Constraints Order Push Pop Split Append Lookup
Vector<A> RRB tree Clone insertion O(1)* O(1)* O(log n) O(log n) O(log n)

Maps

Maps are mappings of keys to values, where the most common read operation is to find the value associated with a given key. Maps may or may not have a defined order. Any given key can only occur once inside a map, and setting a key to a different value will overwrite the previous value.

Type Algorithm Key Constraints Order Insert Remove Lookup
HashMap<K, V> HAMT Clone + Hash + Eq undefined O(log n) O(log n) O(log n)
OrdMap<K, V> B-tree Clone + Ord sorted O(log n) O(log n) O(log n)

Sets

Sets are collections of unique values, and may or may not have a defined order. Their crucial property is that any given value can only exist once in a given set.

Type Algorithm Constraints Order Insert Remove Lookup
HashSet<A> HAMT Clone + Hash + Eq undefined O(log n) O(log n) O(log n)
OrdSet<A> B-tree Clone + Ord sorted O(log n) O(log n) O(log n)

In-place Mutation

All of these data structures support in-place copy-on-write mutation, which means that if you're the sole user of a data structure, you can update it in place without taking the performance hit of making a copy of the data structure before modifying it (this is about an order of magnitude faster than immutable operations, almost as fast as std::collections's mutable data structures).

Thanks to Rc's reference counting, we are able to determine whether a node in a data structure is being shared with other data structures, or whether it's safe to mutate it in place. When it's shared, we'll automatically make a copy of the node before modifying it. The consequence of this is that cloning a data structure becomes a lazy operation: the initial clone is instant, and as you modify the cloned data structure it will clone chunks only where you change them, so that if you change the entire thing you will eventually have performed a full clone.

This also gives us a couple of other optimisations for free: implementations of immutable data structures in other languages often have the idea of local mutation, like Clojure's transients or Haskell's ST monad - a managed scope where you can treat an immutable data structure like a mutable one, gaining a considerable amount of performance because you no longer need to copy your changed nodes for every operation, just the first time you hit a node that's sharing structure. In Rust, we don't need to think about this kind of managed scope, it's all taken care of behind the scenes because of our low level access to the garbage collector (which, in our case, is just a simple Rc).

Re-exports

pub use hashmap::HashMap;
pub use hashset::HashSet;
pub use ordmap::OrdMap;
pub use ordset::OrdSet;
pub use vector::Vector;

Modules

hashmap

An unordered map.

hashset

An unordered set.

iter

Iterators over immutable data.

ordmap

An ordered map.

ordset

An ordered set.

vector

A persistent vector.

Macros

get_in

Get a value inside multiple levels of data structures.

hashmap

Construct a hash map from a sequence of key/value pairs.

hashset

Construct a set from a sequence of values.

ordmap

Construct a map from a sequence of key/value pairs.

ordset

Construct a set from a sequence of values.

update_in

Update a value inside multiple levels of data structures.

vector

Construct a vector from a sequence of elements.