Alkahest is blazing-fast, zero-deps, zero-overhead, zero-unsafe, schema-based serialization library. It is suitable for broad range of use-cases, but tailored for custom high-performance network protocols.
This benchmark that mimics some game networking protocol.
alkahest |
bincode |
rkyv |
speedy |
|
---|---|---|---|---|
serialize |
10.69 us (✅ 1.00x) |
11.08 us (✅ 1.04x slower) |
12.43 us (❌ 1.16x slower) |
11.13 us (✅ 1.04x slower) |
read |
1.19 us (✅ 1.00x) |
9.19 us (❌ 7.74x slower) |
2.10 us (❌ 1.77x slower) |
1.54 us (❌ 1.30x slower) |
Made with criterion-table
See also benchmark results from https://github.com/djkoloski/rust_serialization_benchmark (in draft until 0.2 release).
-
Schema-based serialization. Alkahest uses data schemas called
Formula
s to serialize and deserialize data. Thus controlling data layout independently from data types that are serialized or deserialized. -
Support wide variety of formulas. Integers, floats, booleans, tuples, arrays, slices, strings and user-defined formulas with custom data layout using
derive
macro that works for structs and enums of any complexity and supports generics. -
Zero-overhead serialization of sequences. Alkahest support serializing iterators directly into slice formulas. No more allocation of a
Vec
to serialize and drop immediately. -
Lazy deserialization. Alkahest provides
Lazy<F>
type to deserialize any formulaF
lazily.Lazy
can be used later to perform actual deserialization.
Lazy<[F]>
can also produce iterator that deserializes elements on demand.
Laziness is controlled on type level and can be applied to any element of a larger formula. -
Infallible serialization. Given large enough or growing buffer any value that implements
Serialize
can be serialized without error. No more unnecessary unwraps or puzzles "what to do if serialization fails?". The only error condition for serialization is "data doesn't fit".
- Serializable formula descriptors
- Compatibility rules
- External tool for code-generation for formula descriptors for C and Rust.
Alkahest separates data schema definition (aka Formula
) from
serialization and deserialization code.
Doing so, this library provides better guarantees for cases
when serializable data type and deserializable data type
are different.
It also supports serializing from iterators instead of collections
and deserialization into lazy wrappers that defers costly process
and may omit it entirely if value is never accessed.
User controls laziness on type level by choosing appropriate Deserialize
impls.
For instance deserializing into Vec<T>
is eager because Vec<T>
is constructed
with all T
instances and memory allocated for them.
While alkahest::SliceIter
implements Iterator
and deserializes
elements in Iterator::next
and other methods. And provides constant-time
random access to any element.
Flexibility comes at cost of using only byte slices for serialization and deserialization. And larger footprint of serialized data than some other binary formats.
Question about support of dense data packing is open. It may be desireable to control on type level.
The API is designed with following principles: Any value can be serialized successfully given large enough buffer. Data can't cause panic, incorrect implementation of a trait can.
There is zero unsafe code in the library on any code it generates.
No UB is possible given that std
is not unsound.
No data schemas stays the same. New fields and variants are added, others are deprecated and removed.
There's set of rules that ensures forward compatibility between formulas. And another set or rules for backward compatibility.
Verification of compatibility is not implemented yet.
Forward compatibility is an ability to deserialize data that was serialized with newer formulas.
TODO: List all rules
Backward compatibility is an ability to deserialize data that was serialized with older formulas.
TODO: List all rules
The crate works using three fundamental traits.
Formula
, Serialize
and Deserialize
.
There's also supporting trait - BareFormula
.
Alkahest provides proc-macro alkahest
for deriving Formula
, Serialize
and Deserialize
.
Formula
trait is used to allow types to serve as data schemas.
Any value serialized with given formula should be deserializable with the same
formula. Sharing only Formula
type allows modules and crates
easily communicate.
Formula
dictates binary data layout and it must be platform-independent.
Potentially Formula
types can be generated from separate files,
opening possibility for cross-language communication.
Formula
is implemented for a number of types out-of-the-box.
Primitive types like bool
, integers and floating point types all implement Formula
.
!Caveat!:
Serialized size of isize
and usize
is controlled by a feature-flag.
Sizes and addresses are serialized as usize
.
Truncating usize
value if it was too large.
This may result in broken data generated and panic in debug.
It is also implemented for tuples, array and slice, Option
and Vec
(the later requires "alloc"
feature).
The easiest way to define a new formula is to derive Formula
trait for a struct or an enum.
Generics are supported, but may require complex bounds specified in attributes for
Serialize
and Deserialize
derive macros.
The only constrain is that all fields must implement Formula
.
Serialize<Formula>
trait is used to implement serialization
according to a specific formula.
Serialization writes to mutable bytes slice and should not
perform dynamic allocations.
Binary result of any type serialized with a formula must follow it.
At the end, if a stream of primitives serialized is the same,
binary result should be the same.
Types may be serializable with different formulas producing
different binary result.
Serialize
is implemented for many types.
Most notably there's implementation T: Serialize<T>
and &T: Serialize<T>
for all primitives T
(except usize
and isize
).
Another important implementation is
Serialize<F> for I where I: IntoIterator, I::Item: Serialize<F>
,
allowing serializing into slice directly from both iterators and collections.
Serialization with formula Ref<F>
uses serialization with formula F
and then stores relative address and size. No dynamic allocations is required.
Deriving Serialize
for a type will generate Serialize
implementation,
formula is specified in attribute #[alkahest(FormulaRef)]
or
#[alkahest(serialize(FormulaRef))]
. FormulaRef
is typically a type.
When generics are used it also contains generic parameters and bounds.
If formula is not specified - Self
is assumed.
Formula
should be derived for the type as well.
It is in-advised to derive Serialize
for formulas with
manual Formula
implementation,
Serialize
derive macro generates code that uses non-public items
generated by Formula
derive macro.
So either both should have manual implementation or both derived.
For structures Serialize
derive macro requires that all fields
are present on both Serialize
and Formula
structure and has the same
order (trivially if this is the same structure).
For enums Serialize
derive macro checks that for each variant there
exists variant on Formula
enum.
Variants content is compared similar to structs.
Serialization inserts variant ID and serializes variant as struct.
The size of variants may vary. Padding is inserted by outer value serialization
if necessary.
Serialize
can be derived for structure where Formula
is an enum.
In this case variant should be specified using
#[alkahest(@variant_ident)]
or #[alkahest(serialize(@variant_ident))]
and then Serialize
derive macro will produce serialization code that works
as if this variant was a struct Formula
,
except that variant's ID will be serialized before fields.
Serialize
can be derived for enum only if Formula
is enum as well.
Serializable enum may omit some (or all) variants from Formula
.
It may not have variants missing in Formula
.
Each variant then follows rules for structures.
For convenience Infallible
implements Serialize
for enum formulas.
Deserialize<'de, Formula>
trait is used to implement deserialization
according to a specific formula.
Deserialization reads from bytes slice constructs deserialized value.
Deserialization should not perform dynamic allocations except those
that required to construct and initialize deserialized value.
E.g. it is allowed to allocate when Vec<T>
is produced if non-zero
number of T
values are deserialized. It should not over-allocate.
Similar to Serialize
alkahest provides a number of out-of-the-box
implementations of Deserialize
trait.
From<T>
types can be deserialized with primitive formula T
.
Values that can be deserialized with formula F
can also deserialize with Ref<F>
, it reads address and length
and proceeds with formula F
.
Vec<T>
may deserialize with slice formula.
Deserialize<'de, [F]>
is implemented for alkahest::SliceIter<'de, T>
type
that implements Iterator
and lazily deserialize elements of type
T: Deserialize<'de, F>
. SliceIter
is cloneable,
can be iterated from both ends and skips elements for in constant time.
For convenience SliceIter
also deserializes with array formula.
Deriving Deserialize
for a type will generate Deserialize
implementation,
formula is specified in attribute #[alkahest(FormulaRef)]
or
#[alkahest(deserialize(FormulaRef))]
. FormulaRef
is typically a type.
When generics are used it also contains generic parameters and bounds.
If formula is not specified - Self
is assumed.
Formula
should be derived for the type as well.
It is in-advised to derive Deserialize
for formulas with
manual Formula
implementation,
Deserialize
derive macro generates code that uses non-public items
generated by Formula
derive macro.
So either both should have manual implementation or both derived.
Alkahest is cool but serde
is almost universally used, and for good reasons.
While designing a Formula
it may be desireable to include existing type
that supports serialization serde
, especially if it comes from another crate.
This crate provides Bincode
and Bincoded<T>
formulas to cover this.
Anything with serde::Serialize
implementation can be serialized with Bincode
formula, naturally it will be serialized using bincode
crate.
Bincoded<T>
is a restricted version of Bincode
that works only for T
.
// This requires two default features - "alloc" and "derive".
#[cfg(all(feature = "derive", feature = "alloc"))]
fn main() {
use alkahest::{alkahest, serialize_to_vec, deserialize};
// Define simple formula. Make it self-serializable.
#[derive(Clone, Debug, PartialEq, Eq)]
#[alkahest(Formula, SerializeRef, Deserialize)]
struct MyDataType {
a: u32,
b: Vec<u8>,
}
// Prepare data to serialize.
let value = MyDataType {
a: 1,
b: vec![2, 3],
};
// Use infallible serialization to `Vec`.
let mut data = Vec::new();
// Note that this value can be serialized by reference.
// This is default behavior for `Serialized` derive macro.
// Some types required ownership transfer for serialization.
// Notable example is iterators.
let (size, _) = serialize_to_vec::<MyDataType, _>(&value, &mut data);
let de = deserialize::<MyDataType, MyDataType>(&data[..size]).unwrap();
assert_eq!(de, value);
}
#[cfg(not(all(feature = "derive", feature = "alloc")))]
fn main() {}
Alkahest comes with a benchmark to test against other popular serialization crates.
Simply run cargo bench --all-features
to see results.
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.
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.