Crate fst[][src]

Expand description

Crate fst is a library for efficiently storing and searching ordered sets or maps where the keys are byte strings. A key design goal of this crate is to support storing and searching very large sets or maps (i.e., billions). This means that much effort has gone in to making sure that all operations are memory efficient.

Sets and maps are represented by a finite state machine, which acts as a form of compression on common prefixes and suffixes in the keys. Additionally, finite state machines can be efficiently queried with automata (like regular expressions or Levenshtein distance for fuzzy queries) or lexicographic ranges.

To read more about the mechanics of finite state transducers, including a bibliography for algorithms used in this crate, see the docs for the raw::Fst type.


Simply add a corresponding entry to your Cargo.toml dependency list:

fst = "0.4"

The examples in this documentation will show the rest.

Overview of types and modules

This crate provides the high level abstractions—namely sets and maps—in the top-level module.

The set and map sub-modules contain types specific to sets and maps, such as range queries and streams.

The raw module permits direct interaction with finite state transducers. Namely, the states and transitions of a transducer can be directly accessed with the raw module.

Example: fuzzy query

This example shows how to create a set of strings in memory, and then execute a fuzzy query. Namely, the query looks for all keys within an edit distance of 1 of foo. (Edit distance is the number of character insertions, deletions or substitutions required to get from one string to another. In this case, a character is a Unicode codepoint.)

This requires the levenshtein feature in this crate to be enabled. It is not enabled by default.

use fst::{IntoStreamer, Streamer, Set};
use fst::automaton::Levenshtein;

fn example() -> Result<(), Box<dyn std::error::Error>> {
    // A convenient way to create sets in memory.
    let keys = vec!["fa", "fo", "fob", "focus", "foo", "food", "foul"];
    let set = Set::from_iter(keys)?;

    // Build our fuzzy query.
    let lev = Levenshtein::new("foo", 1)?;

    // Apply our fuzzy query to the set we built.
    let mut stream =;

    let keys = stream.into_strs()?;
    assert_eq!(keys, vec!["fo", "fob", "foo", "food"]);

Warning: Levenshtein automatons use a lot of memory

The construction of Levenshtein automatons should be consider “proof of concept” quality. Namely, they do just enough to be correct. But they haven’t had any effort put into them to be memory conscious.

Note that an error will be returned if a Levenshtein automaton gets too big (tens of MB in heap usage).

Example: stream to a file and memory map it for searching

This shows how to create a MapBuilder that will stream construction of the map to a file. Notably, this will never store the entire transducer in memory. Instead, only constant memory is required during construction.

For the search phase, we use the memmap crate to make the file available as a &[u8] without necessarily reading it all into memory (the operating system will automatically handle that for you).

use std::fs::File;
use std::io;

use fst::{IntoStreamer, Streamer, Map, MapBuilder};
use memmap::Mmap;

// This is where we'll write our map to.
let mut wtr = io::BufWriter::new(File::create("map.fst")?);

// Create a builder that can be used to insert new key-value pairs.
let mut build = MapBuilder::new(wtr)?;
build.insert("bruce", 1).unwrap();
build.insert("clarence", 2).unwrap();
build.insert("stevie", 3).unwrap();

// Finish construction of the map and flush its contents to disk.

// At this point, the map has been constructed. Now we'd like to search it.
// This creates a memory map, which enables searching the map without loading
// all of it into memory.
let mmap = unsafe { Mmap::map(&File::open("map.fst")?)? };
let map = Map::new(mmap)?;

// Query for keys that are greater than or equal to clarence.
let mut stream = map.range().ge("clarence").into_stream();

let kvs = stream.into_str_vec()?;
assert_eq!(kvs, vec![
    ("clarence".to_owned(), 2),
    ("stevie".to_owned(), 3),

Example: case insensitive search

We can perform case insensitive search on a set using a regular expression. We can use the regex-automata crate to compile a regular expression into an automaton:

use fst::{IntoStreamer, Set};
use regex_automata::dense; // regex-automata crate with 'transducer' feature

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let set = Set::from_iter(&["FoO", "Foo", "fOO", "foo"])?;
    let pattern = r"(?i)foo";
    // Setting 'anchored' is important, otherwise the regex can match anywhere
    // in the key. This would cause the regex to iterate over every key in the
    // FST set.
    let dfa = dense::Builder::new().anchored(true).build(pattern).unwrap();

    let keys =;
    assert_eq!(keys, vec!["FoO", "Foo", "fOO", "foo"]);
    println!("{:?}", keys);

Note that for this to work, the regex-automata crate must be compiled with the transducer feature enabled:

fst = "0.4"
regex-automata = { version = "0.1.9", features = ["transducer"] }

Example: searching multiple sets efficiently

Since queries can search a transducer without reading the entire data structure into memory, it is possible to search many transducers very quickly.

This crate provides efficient set/map operations that allow one to combine multiple streams of search results. Each operation only uses memory proportional to the number of streams.

The example below shows how to find all keys that start with B or G. The example below uses sets, but the same operations are available on maps too.

use fst::automaton::{Automaton, Str};
use fst::set;
use fst::{IntoStreamer, Set, Streamer};

fn example() -> Result<(), Box<dyn std::error::Error>> {
    let set1 = Set::from_iter(&["AC/DC", "Aerosmith"])?;
    let set2 = Set::from_iter(&["Bob Seger", "Bruce Springsteen"])?;
    let set3 = Set::from_iter(&["George Thorogood", "Golden Earring"])?;
    let set4 = Set::from_iter(&["Kansas"])?;
    let set5 = Set::from_iter(&["Metallica"])?;

    // Create the matcher. We can reuse it to search all of the sets.
    let matcher = Str::new("B")

    // Build a set operation. All we need to do is add a search result stream
    // for each set and ask for the union. (Other operations, like intersection
    // and difference are also available.)
    let mut stream =

    // Now collect all of the keys. Alternatively, you could build another set
    // here using `SetBuilder::extend_stream`.
    let mut keys = vec![];
    while let Some(key) = {
    assert_eq!(keys, vec![
        "Bob Seger",
        "Bruce Springsteen",
        "George Thorogood",
        "Golden Earring",

Memory usage

An important advantage of using finite state transducers to represent sets and maps is that they can compress very well depending on the distribution of keys. The smaller your set/map is, the more likely it is that it will fit into memory. If it’s in memory, then searching it is faster. Therefore, it is important to do what we can to limit what actually needs to be in memory.

This is where automata shine, because they can be queried in their compressed state without loading the entire data structure into memory. This means that one can store a set/map created by this crate on disk and search it without actually reading the entire set/map into memory. This use case is served well by memory maps, which lets one assign the entire contents of a file to a contiguous region of virtual memory.

Indeed, this crate encourages this mode of operation. Both sets and maps can be constructed from anything that provides an AsRef<[u8]> implementation, which any memory map should.

This is particularly important for long running processes that use this crate, since it enables the operating system to determine which regions of your finite state transducers are actually in memory.

Of course, there are downsides to this approach. Namely, navigating a transducer during a key lookup or a search will likely follow a pattern approximating random access. Supporting random access when reading from disk can be very slow because of how often seek must be called (or, in the case of memory maps, page faults). This is somewhat mitigated by the prevalence of solid state drives where seek time is eliminated. Nevertheless, solid state drives are not ubiquitous and it is possible that the OS will not be smart enough to keep your memory mapped transducers in the page cache. In that case, it is advisable to load the entire transducer into your process’s memory (e.g., calling Set::new with a Vec<u8>).


Searching a set or a map needs to provide some way to iterate over the search results. Idiomatic Rust calls for something satisfying the Iterator trait to be used here. Unfortunately, this is not possible to do efficiently because the Iterator trait does not permit values emitted by the iterator to borrow from the iterator. Borrowing from the iterator is required in our case because keys and values are constructed during iteration.

Namely, if we were to use iterators, then every key would need its own allocation, which could be quite costly.

Instead, this crate provides a Streamer, which can be thought of as a streaming iterator. Namely, a stream in this crate maintains a single key buffer and lends it out on each iteration.

For more details, including important limitations, see the Streamer trait.


There’s no doubt about it, finite state transducers are a specialty data structure. They have a host of restrictions that don’t apply to other similar data structures found in the standard library, such as BTreeSet and BTreeMap. Here are some of them:

  1. Sets can only contain keys that are byte strings.
  2. Maps can also only contain keys that are byte strings, and its values are limited to unsigned 64 bit integers. (The restriction on values may be relaxed some day.)
  3. Creating a set or a map requires inserting keys in lexicographic order. Often, keys are not already sorted, which can make constructing large sets or maps tricky. One way to do it is to sort pieces of the data and build a set/map for each piece. This can be parallelized trivially. Once done, they can be merged together into one big set/map if desired. A somewhat simplistic example of this procedure can be seen in fst-bin/src/ from the root of this crate’s repository.



Automaton implementations for finite state transducers.


Map operations implemented by finite state transducers.


Operations on raw finite state transducers.


Set operations implemented by finite state transducers.



Map is a lexicographically ordered map from byte strings to integers.


A builder for creating a map.


Set is a lexicographically ordered set of byte strings.


A builder for creating a set.



An error that encapsulates all possible errors in this crate.



Automaton describes types that behave as a finite automaton.


IntoStreamer describes types that can be converted to streams.


Streamer describes a “streaming iterator.”

Type Definitions


A Result type alias for this crate’s Error type.