#no-std #serialization

no-std musli

Müsli is a flexible and generic binary serialization framework

49 releases

0.0.49 May 18, 2023
0.0.48 May 18, 2023
0.0.39 Apr 26, 2023
0.0.38 Mar 22, 2023
0.0.29 Apr 30, 2022

#113 in Encoding

Download history 46/week @ 2023-02-07 124/week @ 2023-02-14 182/week @ 2023-02-21 13/week @ 2023-02-28 13/week @ 2023-03-07 15/week @ 2023-03-14 113/week @ 2023-03-21 15/week @ 2023-03-28 5/week @ 2023-04-04 48/week @ 2023-04-11 1/week @ 2023-04-18 45/week @ 2023-04-25 29/week @ 2023-05-02 240/week @ 2023-05-09 360/week @ 2023-05-16 109/week @ 2023-05-23

739 downloads per month
Used in 9 crates


3.5K SLoC


github crates.io docs.rs build status

Excellent performance, no compromises[^1]!

Müsli is a flexible, fast, and generic binary serialization framework for Rust, in the same vein as serde.

It provides a set of formats, each with its own well-documented set of features and tradeoffs.

[^1]: As in Müsli should be able to do everything you need and more.

Quick guide


Add the following to your Cargo.toml using the format you want to use:

musli = "0.0.49"
musli-wire = "0.0.49"


The heavy lifting in user code is done through the Encode and Decode derives which are thoroughly documented in the derives module. Müsli primarily operates based on the schema types which implement these traits imply, but self-descriptive formats are also possible (see Formats below).

use musli::{Encode, Decode};

#[derive(Encode, Decode)]
struct Person {
    /* .. fields .. */

Note by default a field is identified by its numerical index which would change if they are re-ordered. Renaming fields and setting a default naming policy can be done by configuring the derives.

The binary serialization formats provided aim to efficiently and accurately encode every type and data structure available in Rust. Each format comes with well-documented tradeoffs and aim to be fully memory safe to use.

Internally we use the terms "encoding", "encode", and "decode" because it's distinct from serde's use of "serialization", "serialize", and "deserialize" allowing for the ease of using both libraries side by side if desired.

Müsli is designed on similar principles as serde. Relying on Rust's powerful trait system to generate code which can largely be optimized away. The end result should be very similar to handwritten highly optimized code.

As an example of this, these two functions both produce the same assembly on my machine (built with --release):

const ENCODING: Encoding<DefaultMode, Fixed<NativeEndian>, Variable> =

#[derive(Encode, Decode)]
pub struct Storage {
    left: u32,
    right: u32,

fn with_musli(storage: &Storage) -> Result<[u8; 8]> {
    let mut array = [0; 8];
    ENCODING.encode(&mut array[..], storage)?;

fn without_musli(storage: &Storage) -> Result<[u8; 8]> {
    let mut array = [0; 8];

Where Müsli differs in design philosophy is twofold:

We make use of GATs to provide tighter abstractions, which should be easier for Rust to optimize.

We make less use of the Visitor pattern in certain instances where it's deemed unnecessary, such as when decoding collections. The result is usually cleaner decode implementations, as shown here:

use musli::de::{Decode, Decoder, SequenceDecoder};
use musli::mode::Mode;

struct MyType {
    data: Vec<String>,

impl<'de, M> Decode<'de, M> for MyType where M: Mode {
    fn decode<D>(decoder: D) -> Result<Self, D::Error>
        D: Decoder<'de>,
        let mut seq = decoder.decode_sequence()?;
        let mut data = Vec::with_capacity(seq.size_hint().or_default());

        while let Some(decoder) = seq.next()? {


        Ok(Self {

Another major aspect where Müsli differs is in the concept of modes (note the M parameter above). Since this is a parameter of the Encode and Decode traits it allows for the same data model to be serialized in many different ways. This is a larger topic and is covered further down.


Formats are currently distinguished by supporting various degrees of upgrade stability. A fully upgrade stable encoding format must tolerate that one model can add fields that an older version of the model should be capable of ignoring.

Partial upgrade stability can still be useful as is the case of the musli-storage format below, because reading from storage only requires decoding to be upgrade stable. So if correctly managed with #[musli(default)] this will never result in any readers seeing unknown fields.

The available formats and their capabilities are:

reorder missing unknown self
musli-storage #[musli(packed)]

reorder determines whether fields must occur in exactly the order in which they are specified in their type. Reordering fields in such a type would cause unknown but safe behavior of some kind. This is only suitable for byte-oriented IPC where the data models of each client are are strictly synchronized.

missing determines if reading can handle missing fields through something like Option<T>. This is suitable for on-disk storage, because it means that new optional fields can be added as the schema evolves.

unknown determines if the format can skip over unknown fields. This is suitable for network communication. At this point you've reached upgrade stability. Some level of introspection is possible here, because the serialized format must contain enough information about fields to know what to skip which usually allows for reasoning about basic types.

self determines if the format is self-descriptive. Allowing the structure of the data to be fully reconstructed from its serialized state. These formats do not require models to decode, and can be converted to and from dynamic containers such as musli-value for introspection.

For every feature you drop, the format becomes more compact and efficient. musli-storage using #[musli(packed)] for example is roughly as compact as bincode while musli-wire is comparable in size to something like protobuf. All formats are primarily byte-oriented, but some might perform bit packing if the benefits are obvious.

Upgrade stability

The following is an example of full upgrade stability using musli-wire. Note how Version1 can be decoded from an instance of Version2 because it understands how to skip fields which are part of Version2. We're also explicitly #[musli(rename = ..)] the fields to ensure that they don't change in case they are re-ordered.

use musli::{Encode, Decode};

#[derive(Debug, PartialEq, Encode, Decode)]
struct Version1 {
    #[musli(rename = 0)]
    name: String,

#[derive(Debug, PartialEq, Encode, Decode)]
struct Version2 {
    #[musli(rename = 0)]
    name: String,
    #[musli(default, rename = 1)]
    age: Option<u32>,

let version2 = musli_wire::to_buffer(&Version2 {
    name: String::from("Aristotle"),
    age: Some(62),

let version1: Version1 = musli_wire::decode(version2.as_slice())?;

The following is an example of partial upgrade stability using musli-storage on the same data models. Note how Version2 can be decoded from Version1 but not the other way around. That's why it's suitable for on-disk storage the schema can evolve from older to newer versions.

let version2 = musli_storage::to_buffer(&Version2 {
    name: String::from("Aristotle"),
    age: Some(62),

assert!(musli_storage::decode::<_, Version1>(version2.as_slice()).is_err());

let version1 = musli_storage::to_buffer(&Version1 {
    name: String::from("Aristotle"),

let version2: Version2 = musli_storage::decode(version1.as_slice())?;


In Müsli the same model can be serialized in different ways. Instead of requiring the use of multiple models, we instead support each model implementing different modes.

A mode allows for different encoding attributes to apply depending on which mode something is performed in. A mode can apply to any musli parameter giving you a lot of flexibility.

If a mode is not specified, an implementation will apply to all modes (M: Mode), if at least one mode is specified it will be implemented for all modes which are present in a model and DefaultMode. This way, an encoding which uses DefaultMode (which it does by default) should always work.

For more information on how to configure modes, see the derives module. Below is a simple example of how we can use two modes to provide two different kinds of serialization to a single struct.

use musli::mode::{DefaultMode, Mode};
use musli::{Decode, Encode};
use musli_json::Encoding;

enum Alt {}
impl Mode for Alt {}

#[derive(Decode, Encode)]
#[musli(mode = Alt, packed)]
#[musli(default_field_name = "name")]
struct Word<'a> {
    text: &'a str,
    teineigo: bool,

let CONFIG: Encoding<DefaultMode> = Encoding::new();
let ALT_CONFIG: Encoding<Alt> = Encoding::new();

let word = Word {
    text: "あります",
    teineigo: true,

let out = CONFIG.to_string(&word)?;
assert_eq!(out, r#"{"text":"あります","teineigo":true}"#);

let out = ALT_CONFIG.to_string(&word)?;
assert_eq!(out, r#"["あります",true]"#);


This is a non-exhaustive list of unsafe use in this crate, and why they are used:

  • A mem::transcode in Tag::kind. Which guarantees that converting into the Kind enum which is #[repr(u8)] is as efficient as possible.

  • A largely unsafe SliceReader which provides more efficient reading than the default Reader impl for &[u8] does. Since it can perform most of the necessary comparisons directly on the pointers.

  • Some unsafety related to UTF-8 handling in musli_json, because we check UTF-8 validity internally ourselves (like serde_json).

  • FixedBytes<N> is a stack-based container that can operate over uninitialized data. Its implementation is largely unsafe. With it stack-based serialization can be performed which is useful in no-std environments.

  • Some unsafe is used for owned String decoding in all binary formats to support faster string processing using simdutf8. Disabling the simdutf8 feature (enabled by default) removes the use of this unsafe.

To ensure this library is correctly implemented with regards to memory safety, extensive testing is performed using miri. For more information on this, see musli-tests for more information on this.


The following are the results of preliminary benchmarking and should be taken with a big grain of 🧂.

The two benchmark suites portrayed are:

  • rt-prim - which is a small object containing one of each primitive type and a string and a byte array.
  • rt-lg - which is roundtrip encoding of a large object, containing vectors and maps of other objects.
Roundtrip of a large object Roundtrip of a small object

Size comparisons

This is not yet an area which has received much focus, but because people are bound to ask the following section performs a raw size comparison between different formats.

Each test suite serializes a collection of values, which have all been randomly populated.

  • A struct containing one of every primitive value (prim).
  • A really big struct (lg).
  • A structure containing fairly sizable, allocated fields (allocated).
  • A moderately sized enum with many field variations (medium_enum).

Note so far these are all synthetic examples. Real world data is rarely this random. But hopefully it should give an idea of the extreme ranges.

framework prim lg allocated medium_enum
derive_bitcode[^i128] 54.94 ± 0.25 9442.00 ± 2786.37 552.50 ± 312.94 104.28 ± 223.41
musli_descriptive 95.06 ± 1.31 12995.40 ± 3802.46 636.98 ± 317.05 111.17 ± 222.81
musli_json[^incomplete] 180.78 ± 2.29 19913.10 ± 6081.87 703.58 ± 322.02 133.13 ± 225.23
musli_storage 61.74 ± 0.44 10433.10 ± 3070.87 547.48 ± 312.98 101.10 ± 222.83
musli_storage_packed 48.74 ± 0.44 9604.90 ± 2811.17 544.48 ± 312.98 99.26 ± 223.07
musli_wire 81.50 ± 1.40 12041.30 ± 3470.11 592.87 ± 314.66 107.70 ± 222.80
rkyv[^incomplete] 56.00 ± 0.00 13239.20 ± 3425.15 562.04 ± 312.82 158.03 ± 226.79
serde_bincode 54.94 ± 0.25 9821.70 ± 2853.71 564.66 ± 312.73 108.52 ± 225.24
serde_bitcode[^i128] 54.94 ± 0.25 9449.50 ± 2786.20 552.50 ± 312.94 104.27 ± 223.40
serde_cbor[^i128] 174.08 ± 0.79 18810.50 ± 5987.69 612.80 ± 315.17 136.06 ± 224.74
serde_dlhn[^i128] 55.05 ± 0.92 10078.10 ± 2926.49 544.48 ± 312.98 99.53 ± 223.13
serde_json[^incomplete] 266.78 ± 2.29 26306.80 ± 8600.79 712.58 ± 322.02 159.36 ± 232.86
serde_rmp 61.02 ± 1.08 11477.60 ± 3136.97 577.12 ± 314.19 117.70 ± 222.53

[^incomplete]: These formats do not support a wide range of Rust types. Exact level of support varies. But from a size perspective it makes size comparisons either unfair or simply an esoteric exercise since they can (or cannot) make stricter assumptions as a result.

[^i128]: Lacks 128-bit support.


~22K SLoC