revision 0.17.0

Information

Revision is a framework for revision-tolerant serialization and deserialization with support for schema evolution over time. It allows for easy revisioning of structs and enums for data storage requirements which need to support backwards compatibility, but where the design of the data structures evolve over time. Revision enables data that was serialized at older revisions to be seamlessly deserialized and converted into the latest data structures. It uses bincode for serialization and deserialization.

The Revisioned trait is automatically implemented for the following primitives: u8, u16, u32, u64, u128, usize, i8, i16, i32, i64, i128, isize, f32, f64, char, String, Vec<T>, Arrays up to 32 elements, Option<T>, Box<T>, Bound<T>, Wrapping<T>, Reverse<T>, (A, B), (A, B, C), (A, B, C, D), (A, B, C, D, E), Duration, HashMap<K, V>, BTreeMap<K, V>, HashSet<T>, BTreeSet<T>, BinaryHeap<T>, Result<T, E>, Cow<'_, T>, Decimal, regex::Regex, uuid::Uuid, chrono::Duration, chrono::DateTime<Utc>, geo::Point, geo::LineString geo::Polygon, geo::MultiPoint, geo::MultiLineString, geo::MultiPolygon, and ordered_float::NotNan.

Feature Flags

Revision supports the following feature flags:

• specialised-vectors (default): Enables specialised implementations for certain vector types that provide serialisation and deserialisation performance improvements.
• fixed-width-encoding: Uses fixed-width encoding for integers instead of variable-length encoding. By default, Revision uses variable-length encoding which is more space-efficient for small values but has overhead for large values. With this feature enabled, all integers use their full size (2 bytes for u16/i16, 4 bytes for u32/i32, 8 bytes for u64/i64, 16 bytes for u128/i128), providing predictable serialization sizes, and improved serialisation and deserialisation performance.

Integer Encoding Trade-offs

Variable-length encoding (default):

• Small values (0-250) use only 1 byte
• More compact for typical workloads with mostly small values
• Variable serialization size based on value magnitude
• Slight overhead for very large values

Fixed-width encoding (fixed-width-encoding feature):

• Predictable, constant serialization size per type
• No branching or size checks during encoding/decoding
• Less compact for small values
• More efficient for workloads with large values

Benchmarking

To compare variable-length vs fixed-width encoding performance:

# Benchmark with default variable-length encoding
cargo bench --bench varint_comparison

# Benchmark with fixed-width encoding
cargo bench --bench varint_comparison --features fixed-width-encoding

The varint_comparison benchmark tests serialization and deserialization performance across different data distributions (small values, large values, and mixed distributions) for all integer types.

Inspiration

This code takes inspiration from the Versionize library developed for Amazon Firecracker snapshot-restore development previews.

Revision in action

use revision::Error;
use revision::revisioned;

// The test structure is at revision 3.
#[revisioned(revision = 3)]
#[derive(Debug, PartialEq)]
pub struct TestStruct {
    a: u32,
    #[revision(start = 2, end = 3, convert_fn = "convert_b")]
    b: u8,
    #[revision(start = 3)]
    c: u64,
    #[revision(start = 3, default_fn = "default_c")]
    d: String,
}

impl TestStruct {
    // Used to set the default value for a newly added field.
    fn default_c(_revision: u16) -> Result<String, Error> {
        Ok("test_string".to_owned())
    }
    // Used to convert the field from an old revision to the latest revision
    fn convert_b(&mut self, _revision: u16, value: u8) -> Result<(), Error> {
        self.c = value as u64;
        Ok(())
    }
}

// The test structure is at revision 3.
#[revisioned(revision = 3)]
#[derive(Debug, PartialEq)]
pub enum TestEnum {
    #[revision(end = 2, convert_fn = "upgrade_zero")]
    Zero,
    #[revision(end = 2, convert_fn = "upgrade_one")]
    One(u32),
    #[revision(start = 2)]
    Two(u64),
    #[revision(start = 2)]
    Three {
        a: i64,
        #[revision(end = 3, convert_fn = "upgrade_three_b")]
        b: f32,
        #[revision(start = 2)]
        c: rust_decimal::Decimal,
        #[revision(start = 3)]
        d: String,
    },
}

impl TestEnum {
    // Used to convert an old enum variant into a new variant.
    fn upgrade_zero(_: TestEnumZeroFields, _revision: u16) -> Result<TestEnum, Error> {
        Ok(Self::Two(0))
    }
    // Used to convert an old enum variant into a new variant.
    fn upgrade_one(f: TestEnumOneFields, _revision: u16) -> Result<TestEnum, Error> {
        Ok(Self::Two(f.0 as u64))
    }
    // Used to convert the field from an old revision to the latest revision
    fn upgrade_three_b(
        res: &mut TestEnumThreeFields,
        _revision: u16,
        value: f32,
    ) -> Result<(), Error> {
        res.c = value.into();
        Ok(())
    }
}