Serde Rust Serialization Framework

Serde is a powerful framework that enables serialization libraries to generically serialize Rust data structures without the overhead of runtime type information. In many situations, the handshake protocol between serializers and serializees can be completely optimized away, leaving Serde to perform roughly the same speed as a hand written serializer for a specific type.

Documentation is available at http://erickt.github.io/rust-serde/serde

Making a Type Serializable

The simplest way to make a type serializable is to use the serde_macros syntax extension, which comes with a #[derive(Serialize, Deserialize)] annotation, which automatically generates implementations of Serialize and Deserialize for the annotated type:

#[feature(custom_derive, plugin)]
#[plugin(serde_macros)]

extern crate serde;

...

#[derive(Serialize, Deserialize)]
struct Point {
    x: i32,
    y: i32,
}

Serde bundles a high performance JSON serializer and deserializer, serde::json, which comes with the helper functions to_string and from_str that make it easy to go to and from JSON:

use serde::json;

...

let point = Point { x: 1, y: 2 };
let serialized_point = json::to_string(&point).unwrap();

println!("{}", serialized_point); // prints: {"x":1,"y":2}

let deserialize_point: Point = json::from_str(&serialized_point).unwrap();

serde::json also supports a generic Value type, which can represent any JSON value. Also, any Serialize and Deserialize can be converted into a Value with the methods to_value and from_value:

let point = Point { x: 1, y: 2 };
let point_value = json::to_value(&point).unwrap();

println!("{}", point_value.find("x")); // prints: Some(1)

let deserialize_point: Point = json::from_value(point_value).unwrap();

Serialization without Macros

Under the covers, Serde extensively uses the Visitor pattern to thread state between the Serializer and Serialize without the two having specific information about each other's concrete type. This has many of the same benefits as frameworks that use runtime type information without the overhead. In fact, when compiling with optimizations, Rust is able to remove most or all the visitor state, and generate code that's nearly as fast as a hand written serializer format for a specific type.

To see it in action, lets look at how a simple type like i32 is serialized. The Serializer is threaded through the type:

impl serde::Serialize for i32 {
    fn serialize<S>(&self, serializer: &mut S) -> Result<(), S::Error>
        where S: serde::Serializer,
    {
        serializer.visit_i32(*self)
    }
}

As you can see it's pretty simple. More complex types like BTreeMap need to pass a MapVisitor to the Serializer in order to walk through the type:

impl<K, V> Serialize for BTreeMap<K, V>
    where K: Serialize + Ord,
          V: Serialize,
{
    #[inline]
    fn serialize<S>(&self, serializer: &mut S) -> Result<(), S::Error>
        where S: Serializer,
    {
        serializer.visit_map(MapIteratorVisitor::new(self.iter(), Some(self.len())))
    }
}

pub struct MapIteratorVisitor<Iter> {
    iter: Iter,
    len: Option<usize>,
}

impl<K, V, Iter> MapIteratorVisitor<Iter>
    where Iter: Iterator<Item=(K, V)>
{
    #[inline]
    pub fn new(iter: Iter, len: Option<usize>) -> MapIteratorVisitor<Iter> {
        MapIteratorVisitor {
            iter: iter,
            len: len,
        }
    }
}

impl<K, V, I> MapVisitor for MapIteratorVisitor<I>
    where K: Serialize,
          V: Serialize,
          I: Iterator<Item=(K, V)>,
{
    #[inline]
    fn visit<S>(&mut self, serializer: &mut S) -> Result<Option<()>, S::Error>
        where S: Serializer,
    {
        match self.iter.next() {
            Some((key, value)) => {
                let value = try!(serializer.visit_map_elt(key, value));
                Ok(Some(value))
            }
            None => Ok(None)
        }
    }

    #[inline]
    fn len(&self) -> Option<usize> {
        self.len
    }
}

Serializing structs follow this same pattern. In fact, structs are represented as a named map. It's visitor uses a simple state machine to iterate through all the fields:

struct Point {
    x: i32,
    y: i32,
}

impl serde::Serialize for Point {
    fn serialize<S>(&self, serializer: &mut S) -> Result<(), S::Error>
        where S: serde::Serializer
    {
        serializer.visit_named_map("Point", PointMapVisitor {
            value: self,
            state: 0,
        })
    }
}

struct PointMapVisitor<'a> {
    value: &'a Point,
    state: u8,
}

impl<'a> serde::ser::MapVisitor for PointMapVisitor {
    fn visit<S>(&mut self, serializer: &mut S) -> Result<Option(), S::Error>
        where S: serde::Serializer
    {
        match self.state {
            0 => {
                self.state += 1;
                Ok(Some(try!(serializer.visit_map_elt("x", &self.x)))
            }
            1 => {
                self.state += 1;
                Ok(Some(try!(serializer.visit_map_elt("y", &self.y))))
            }
            _ => {
                Ok(None)
            }
        }
    }
}

Deserialization without Macros

Deserialization is a little more complicated since there's a bit more error handling that needs to occur. Let's start with the simple i32 Deserialize implementation. It passes a Visitor to the Deserializer. The Visitor can create the i32 from a variety of different types:

impl Deserialize for i32 {
    fn deserialize<D>(deserializer: &mut D) -> Result<$ty, D::Error>
        where D: serde::Deserializer,
    {
        deserializer.visit(I32Visitor)
    }
}

struct I32Visitor;

impl serde::de::Visitor for I32Visitor {
    type Value = i32;

    fn visit_i16<E>(&mut self, value: i16) -> Result<i16, E>
        where E: Error,
    {
        self.visit_i32(value as i32)
    }

    fn visit_i32<E>(&mut self, value: i32) -> Result<i32, E>
        where E: Error,
    {
        Ok(value)
    }

    ...

Since it's possible for this type to get passed an unexpected type, we need a way to error out. This is done by way of the Error trait, which allows a Deserialize to generate an error for a few common error conditions. Here's how it could be used:

    ...

    fn visit_string<E>(&mut self, _: String) -> Result<i32, E>
        where E: Error,
    {
        Err(serde::de::Error::syntax_error())
    }

    ...

Maps follow a similar pattern as before, and use a MapVisitor to walk through the values generated by the Deserializer.

impl<K, V> serde::Deserialize for BTreeMap<K, V>
    where K: serde::Deserialize + Eq + Ord,
          V: serde::Deserialize,
{
    fn deserialize<D>(deserializer: &mut D) -> Result<BTreeMap<K, V>, D::Error>
        where D: serde::Deserializer,
    {
        deserializer.visit(BTreeMapVisitor::new())
    }
}

pub struct BTreeMapVisitor<K, V> {
    marker: PhantomData<BTreeMap<K, V>>,
}

impl<K, V> BTreeMapVisitor<K, V> {
    pub fn new() -> Self {
        BTreeMapVisitor {
            marker: PhantomData,
        }
    }
}

impl<K, V> serde::de::Visitor for BTreeMapVisitor<K, V>
    where K: serde::de::Deserialize + Ord,
          V: serde::de::Deserialize
{
    type Value = BTreeMap<K, V>;

    fn visit_unit<E>(&mut self) -> Result<BTreeMap<K, V>, E>
        where E: Error,
    {
        Ok(BTreeMap::new())
    }

    fn visit_map<V_>(&mut self, mut visitor: V_) -> Result<BTreeMap<K, V>, V_::Error>
        where V_: MapVisitor,
    {
        let mut values = BTreeMap::new();

        while let Some((key, value)) = try!(visitor.visit()) {
            values.insert(key, value);
        }

        try!(visitor.end());

        Ok(values)
    }
}

Deserializing structs goes a step further in order to support not allocating a String to hold the field names. This is done by custom field enum that deserializes an enum variant from a string. So for our Point example from before, we need to generate:

enum PointField {
    X,
    Y,
}

impl serde::Deserialize for PointField {
    fn deserialize<D>(deserializer: &mut D) -> Result<PointField, D::Error>
        where D: serde::de::Deserializer
    {
        struct FieldVisitor;

        impl serde::de::Visitor for FieldVisitor {
            type Value = Field;

            fn visit_str<E>(&mut self, value: &str) -> Result<PointField, E>
                where E: serde::de::Error
            {
                match value {
                    "x" => Ok(Field::X),
                    "y" => Ok(Field::Y),
                    _ => Err(serde::de::Error::syntax_error()),
                }
            }
        }

        deserializer.visit(FieldVisitor)
    }
}

This is then used in our actual deserializer:

impl serde::Deserialize for Point {
    fn deserialize<D>(deserializer: &mut D) -> Result<Point, D::Error>
        where D: serde::de::Deserializer
    {
        deserializer.visit_named_map("Point", PointVisitor)
    }
}

struct PointVisitor;

impl serde::de::Visitor for PointVisitor {
    type Value = Point;

    fn visit_map<V>(&mut self, mut visitor: V) -> Result<Point, V::Error>
        where V: serde::de::MapVisitor
    {
        let mut x = None;
        let mut y = None;

        loop {
            match try!(visitor.visit_key()) {
                Some(Field::X) => { x = Some(try!(visitor.visit_value())); }
                Some(Field::Y) => { y = Some(try!(visitor.visit_value())); }
                None => { break; }
            }
        }

        let x = match x {
            Some(x) => x,
            None => try!(visitor.missing_field("x")),
        };

        let y = match y {
            Some(y) => y,
            None => try!(visitor.missing_field("y")),
        };

        try!(visitor.end());

        Ok(Point{ x: x, y: y })
    }
}