Lib.rs

A framework for revision-tolerant serialization and deserialization, with support for schema evolution over time, allowing for easy revisioning of structs and enums for data storage requirements which need to support backwards compatibility, but where the design of the data format evolves over time.

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.
• skip (disabled by default): Enables SkipRevisioned / SkipCheckRevisioned, skip_slice / skip_check_slice (plus skip_reader / skip_check_reader aliases), slice fast paths, and matching derive output (#[revisioned(..., skip = false)] opts out per type). Library crates should forward skip = ["revision/skip"] and document features = ["skip"] for dependents; see Skipping encoded values below.

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(())
    }
}

Skipping encoded values

Use the skip feature when you handle revisioned bytes but only need to extract certain fields from the binary data - without deserializing full structs or maps into memory.

Extracting one field from a struct

A #[revisioned] struct is laid out as struct revision (u16), then fields in source order. Read only what you need and call SkipRevisioned::skip_revisioned on &mut reader for the rest (or use skip_slice::<T> to skip a whole nested value in one go when you have a sub-slice).

use revision::{DeserializeRevisioned, Error, SkipRevisioned, revisioned, to_vec};

#[revisioned(revision = 1)]
struct Row {
    // Large field we do not want to allocate when we only need `id`.
    blob: Vec<u8>,
    id: u64,
}

fn read_row_id_only(mut reader: &[u8]) -> Result<u64, Error> {
    let _struct_revision = u16::deserialize_revisioned(&mut reader)?;
    <Vec<u8> as SkipRevisioned>::skip_revisioned(&mut reader)?;
    u64::deserialize_revisioned(&mut reader)
}

let row = Row {
    blob: vec![1, 2, 3],
    id: 42,
};
let bytes = to_vec(&row).unwrap();
assert_eq!(read_row_id_only(&bytes).unwrap(), 42);

Extracting one entry from a BTreeMap

Maps are encoded as length (usize), then key / value pairs in sorted key order. Typical pattern: deserialize each key, compare, deserialize the value you care about, otherwise skip the value with the appropriate skip_revisioned call.

use revision::{DeserializeRevisioned, Error, SkipRevisioned, revisioned, to_vec};
use std::collections::BTreeMap;

#[revisioned(revision = 1)]
struct Config {
    values: BTreeMap<String, u64>,
}

fn get_u64(mut reader: &[u8], wanted: &str) -> Result<u64, Error> {
    let _struct_revision = u16::deserialize_revisioned(&mut reader)?;
    let n = usize::deserialize_revisioned(&mut reader)?;
    for _ in 0..n {
        let key = String::deserialize_revisioned(&mut reader)?;
        if key == wanted {
            return u64::deserialize_revisioned(&mut reader);
        }
        <u64 as SkipRevisioned>::skip_revisioned(&mut reader)?;
    }
    Err(Error::Deserialize(format!("missing key `{wanted}`")))
}

let cfg = Config {
    values: BTreeMap::from([
        ("noise".into(), 0),
        ("answer".into(), 99),
    ]),
};
let bytes = to_vec(&cfg).unwrap();
assert_eq!(get_u64(&bytes, "answer").unwrap(), 99);

For map values that are themselves #[revisioned] enums or structs, deserialize the discriminant / nested revision as you would when fully deserializing, and call MyValue::skip_revisioned on entries you discard (see benches/skip_mixed_btreemap_nested.rs).

Use skip_check_* when you want validation that matches stricter deserialize checks (e.g. UTF-8 for String). Disable skip for a type with #[revisioned(revision = N, skip = false)].

Walking encoded values

WalkRevisioned is a higher-level companion to SkipRevisioned: it lets a caller progress element-by-element through revisioned bytes, deciding per-element whether to decode, skip, or walk into further structure — without rewriting the byte-arithmetic by hand each time. The trait sits between DeserializeRevisioned (decode the entire value) and SkipRevisioned (consume the whole encoding).

The derive macro emits WalkRevisioned for every #[revisioned(...)] type by default (controlled by the same flag as deserialize). Opt out per type with #[revisioned(revision = N, walk = false)].

For each #[revisioned(...)] type the derive emits a per-type walker (<TypeName>Walker<'r, R>) with named per-field / per-variant methods. This is in addition to the generic StructWalker / EnumWalker / MapWalker / SeqWalker types that hand-written WalkRevisioned impls can return.

Walking a struct

use revision::{WalkRevisioned, revisioned, to_vec};

#[revisioned(revision = 1)]
struct Row {
    blob: Vec<u8>,
    id: u64,
}

fn read_row_id_only(mut reader: &[u8]) -> Result<u64, revision::Error> {
    let mut walker = Row::walk_revisioned(&mut reader)?;
    walker.skip_blob()?;
    walker.decode_id()
}

Walking a map

BTreeMap<K, V> returns a MapWalker whose next_entry borrows one key/value pair at a time. Decode the key, then either decode/skip/walk the value before moving on:

use revision::{MapWalker, WalkRevisioned, to_vec};
use std::collections::BTreeMap;

let mut map: BTreeMap<String, u64> = BTreeMap::new();
map.insert("noise".into(), 0);
map.insert("answer".into(), 99);
let bytes = to_vec(&map).unwrap();

let mut reader = bytes.as_slice();
let mut walker: MapWalker<String, u64, _> = <BTreeMap<String, u64>>::walk_revisioned(&mut reader)?;
let mut found = None;
while let Some(mut entry) = walker.next_entry() {
    let k = entry.decode_key()?;
    if k == "answer" {
        found = Some(entry.decode_value()?);
    } else {
        entry.skip_value()?;
    }
}
assert_eq!(found, Some(99));

Walking an enum

For each variant, the derive emits an into_<variant> consuming method that descends into the variant's payload (for unit and single-field tuple variants), and a per-revision walk_revisioned_variant_name(wire_rev, disc) lookup:

use revision::{WalkRevisioned, revisioned, to_vec};

#[revisioned(revision = 1)]
#[derive(Debug, PartialEq)]
enum Shape {
    Square(u32),
    Rectangle { w: u32, h: u32 },
    Circle(u32),
}

let bytes = to_vec(&Shape::Circle(7)).unwrap();
let mut reader = bytes.as_slice();
let walker = Shape::walk_revisioned(&mut reader)?;
if walker.is_circle() {
    let inner = walker.into_circle()?;
    let radius = inner.decode()?;
    assert_eq!(radius, 7);
}

Walking across revisions

WalkRevisioned honours the same cross-revision contract as DeserializeRevisioned: any wire revision in 1..=current is accepted, and the walker presents the latest schema view. The walker repr has up to four arms depending on the type:

• Wire (the fast path) is used when the wire revision matches the current schema, and for any older revision of a type that does not use convert_fn. Per-field methods branch on wire_rev against the field's start annotation: fields added after the wire revision are synthesised via Default::default() (or the user-supplied default_fn); no allocations.
• IndexedBorrowed (struct walker only) holds a borrowed slice over an optimised + indexed_struct payload. Per-field methods jump via the offset table in O(1); no allocations.
• OptimisedBorrowed (enum walker only) holds a borrowed slice over an optimised enum's variant body. Per-variant accessors read directly from the slice.
• ConvertedOwned is used when the wire revision differs from the current schema and the type has at least one convert_fn. The walker internally calls Self::deserialize_revisioned (which honours convert_fn), re-encodes the result at the current revision into an owned Vec<u8>, and then byte-walks those new bytes. The user-facing API is identical; the cost is a single Vec<u8> allocation plus the deserialize/serialize roundtrip.

The walker's repr is selected at construction; per-method code paths do not branch beyond a single match on the internal repr.

use revision::{WalkRevisioned, revisioned, to_vec};

#[revisioned(revision = 1)]
struct ShapeV1 {
    kind: u8,
}

#[revisioned(revision = 2)]
struct Shape {
    kind: u8,
    #[revision(start = 2)]
    flags: u8,
}

let bytes = to_vec(&ShapeV1 { kind: 3 }).unwrap();
let mut r = bytes.as_slice();
let mut walker = Shape::walk_revisioned(&mut r)?;
let kind = walker.decode_kind()?;   // exists at all revisions
let flags = walker.decode_flags()?; // synthesised default at wire rev 1
assert_eq!((kind, flags), (3, 0));

Performance characteristics

Zero-copy peeking

When a walker visits a value whose wire format is usize len || raw bytes — a string, a Vec<u8>, a PathBuf, or any newtype wrapping one — the caller usually wants to compare those bytes against a needle, hash them, or stream them somewhere. Decoding the value just to throw the owned String / Vec<u8> / Bytes away is pure overhead.

Two small traits unlock zero-copy peeking on those payloads:

When both are satisfied, walkers expose the following methods:

Worked example: matching a map key by raw bytes

MapWalker::find_bytes is the direct analogue of find, but the predicate sees the key's wire bytes instead of a decoded K:

use std::collections::BTreeMap;
use revision::{MapWalker, WalkRevisioned, to_vec};

let mut table = BTreeMap::new();
table.insert("alpha".to_string(), 1u32);
table.insert("delta".to_string(), 2);
table.insert("zeta".to_string(), 3);
let bytes = to_vec(&table).unwrap();

let mut r = bytes.as_slice();
let walker: MapWalker<String, u32, _> =
    <BTreeMap<String, u32>>::walk_revisioned(&mut r).unwrap();

// Compare keys as `&[u8]` — no Strand / String allocated per visit.
let value = walker
    .find_bytes(|k| k.cmp(b"delta".as_slice()))
    .unwrap()
    .map(|leaf| leaf.decode())
    .transpose()
    .unwrap();

assert_eq!(value, Some(2));

Worked example: peeking a single key during streaming iteration

MapEntry::with_key_bytes is the per-entry counterpart. Use it when iterating with next_entry and you want to decide what to do with the value based on the key's bytes:

use std::collections::BTreeMap;
use revision::{MapWalker, WalkRevisioned, to_vec};

let mut table = BTreeMap::new();
table.insert("alpha".to_string(), 1u32);
table.insert("beta".to_string(), 2);
table.insert("gamma".to_string(), 3);
let bytes = to_vec(&table).unwrap();

let mut r = bytes.as_slice();
let mut walker: MapWalker<String, u32, _> =
    <BTreeMap<String, u32>>::walk_revisioned(&mut r).unwrap();

let mut beta = None;
while let Some(mut entry) = walker.next_entry() {
    let is_target = entry.with_key_bytes(|k| k == b"beta").unwrap();
    if is_target {
        beta = Some(entry.decode_value().unwrap());
    } else {
        entry.skip_value().unwrap();
    }
}
assert_eq!(beta, Some(2));

Worked example: filtering a map by value bytes

MapEntry::with_value_bytes mirrors with_key_bytes for the value slot. Useful when the key has already been handled (decoded or skipped) and the caller wants to filter based on the value's raw bytes:

use std::collections::BTreeMap;
use revision::{MapWalker, WalkRevisioned, to_vec};

let mut table: BTreeMap<String, Vec<u8>> = BTreeMap::new();
table.insert("a".into(), b"first-value".to_vec());
table.insert("b".into(), b"target-value".to_vec());
let bytes = to_vec(&table).unwrap();

let mut r = bytes.as_slice();
let mut walker: MapWalker<String, Vec<u8>, _> =
    <BTreeMap<String, Vec<u8>>>::walk_revisioned(&mut r).unwrap();

let mut hits = 0;
while let Some(mut entry) = walker.next_entry() {
    entry.skip_key().unwrap();
    if entry.with_value_bytes(|raw| raw.starts_with(b"target")).unwrap() {
        hits += 1;
    }
}
assert_eq!(hits, 1);

Worked example: scanning a sequence of strings

SeqItem::with_bytes lets a scan over Vec<String> (or any SeqWalker whose item type implements LengthPrefixedBytes) compare items as raw bytes without paying for a per-item allocation:

use revision::{SeqWalker, WalkRevisioned, to_vec};

let v = vec!["alpha".to_string(), "beta".into(), "gamma".into()];
let bytes = to_vec(&v).unwrap();

let mut r = bytes.as_slice();
let mut walker: SeqWalker<String, _> =
    <Vec<String>>::walk_revisioned(&mut r).unwrap();

let mut found = false;
while let Some(item) = walker.next_item() {
    if item.with_bytes(|s| s == b"beta").unwrap() {
        found = true;
    }
}
assert!(found);

When zero-copy peeking does not apply

• The reader is a streaming source (std::fs::File, TcpStream, …). BorrowedReader is only implemented for slice-backed readers.
• The element type is a derived #[revisioned(...)] type. Its wire format includes a u16 revision header followed by the body, not bare length-prefixed bytes; use decode / walk and let the walker read past the header.
• The element is a primitive numeric (u32, f64, …) or a fixed-size array. There is no length prefix; the wire bytes are the value bytes. Use decode directly.

Limitations

• Untrusted inputs: Wire lengths are usize length prefixes like everywhere else in revision; they bound how much is read, skipped, or materialised. Walkers add no extra caps or validation — same trust model as DeserializeRevisioned / SkipRevisioned.

• MapWalker::find / find_bytes: On a match you only get a LeafWalker for that entry's value. The method consumes the MapWalker; you cannot resume next_entry on it. Key–value pairs that sort after the match remain on the underlying reader for other callers, not for the same walker instance (by design). Both methods assume wire visit order matches sorted-map encoding (as when serialising BTreeMap). Using an ordering predicate on bytes produced from unsorted maps (HashMap insertion order, …) can match incorrectly or discard the tail under Ordering::Greater.

• LengthPrefixedBytes on custom types: The marker must match the type's real SerializeRevisioned layout (usize len || raw bytes). A wrong impl breaks with_bytes / find_bytes and related paths — it is an explicit contract, not something the library can detect (same class of risk as any incorrect Revisioned impl).

• The derive emits two flavours of nested walk per field. walk_<field>(&mut self) borrows the parent walker so the caller can keep reading siblings after the sub-walker is dropped. into_walk_<field>(self) consumes the parent and hands the reader to the sub-walker for the original 'r, trading sibling access for a longer-lived sub-walker. Both error with Error::Conversion on the ConvertedOwned repr (older revs of convert_fn-bearing types); callers that hit that path should decode_<field> instead.

• into_<variant> is currently emitted for unit variants and single-field tuple variants. Multi-field tuple variants and struct variants are reachable via discriminant() + decode_<field> on the underlying bytes.

• Vec<T> uses specialised-vectors bulk encoding for several element types when that Cargo feature is enabled (the default): primitives, bool, and — if the optional uuid / rust_decimal crate features are also enabled — uuid::Uuid and rust_decimal::Decimal (see try_specialized! in src/implementations/vecs.rs). Vec<T>::walk_revisioned rejects each such T with Error::Deserialize before reading the sequence length, leaving the reader unchanged — use DeserializeRevisioned or SkipRevisioned instead. With specialised-vectors disabled, every Vec<T> uses per-element layout and is safe to walk. HashSet<T>, BTreeSet<T>, BinaryHeap<T>, and the imbl collections always use per-element framing, so they are walkable regardless of element type.

• MapEntry methods enforce key/value ordering in every build: calling decode_value before decode_key / skip_key, or repeating decode_key, returns Error::Deserialize without advancing the reader when the check fails before I/O.

• SeqItem::walk, MapEntry::walk_value, and StructWalker::walk advance counters (remaining, position) only after walk_revisioned succeeds, so a failed nested walk does not desynchronise the parent walker from the byte stream.

• A type using convert_fn requires both serialize = true and deserialize = true for walk to be derivable (the default). The derive errors at compile time if walk = true is combined with either disabled, since the ConvertedOwned cross-revision path needs to deserialize at the wire revision and re-serialize at the current revision. Set walk = false on such a type if you don't need walker support.

• Cow<'_, T> is treated as opaque by the walker. Its Walker is a LeafWalker<T::Owned>, so decode() returns T::Owned (e.g. String for Cow<'_, str>), not a Cow. Use DeserializeRevisioned if you need a Cow back, or descend through T::Owned::walk_revisioned directly.

Optimised wire format

revision 0.23 introduces an opt-in optimised wire format that trades the default varint+sequential layout for a more compact tagged envelope with O(1) skip and optional O(1)/O(log n) random access. Types declare which revisions use it via the history syntax:

#[revisioned(
    revision(1),                                      // legacy layout
    revision(2, optimised),              // tagged envelope
    revision(3, optimised, indexed_struct),
)]
struct Wide { /* fields */ }

Legacy #[revisioned(revision = N)] keeps working — it is normalised internally to revision(1), revision(2), ..., revision(N) all-legacy. The parser distinguishes the two by peeking the next token after the revision keyword (= for legacy, ( for the new function-call form).

History semantics

• Revisions are strict-append. Numbers must run 1..=N with no gaps and no duplicates; the parser errors at the call site otherwise.
• Mixing revision = N with revision(N) on the same type is a compile error.
• Encoding-specific attributes (indexed_struct and the per-field indexed_map / indexed_seq / indexed_set markers) require the optimised flag on the same revision entry.

Wire layout (per-entry)

A type's outer envelope still begins with the u16 revision varint. Under optimised the body that follows is:

struct:  u32_le payload_length || [optional u32_le; field_count] || fields
enum:    u8 tag                || payload per size class

The 1-byte enum tag packs the variant id (bits 0..=4) with a size class (bits 5..=6):

Every variant of an optimised enum must declare its size class via #[revision(size = "inline" | "fixed(N)" | "varlen")]. Variant id is the existing CalcDiscriminant output validated to fit in 5 bits; optimised enums may have at most 32 variants alive at any revision.

Indexed prologues

indexed_struct prepends [u32_le; field_count] to the payload so a walker can jump to any field in O(1). The encoder buffers fields into a scratch Vec<u8> to learn each field's offset, then emits the prologue and body in a single pass. Indexed encoding for individual map/seq/set fields uses the per-field attributes #[revision(indexed_map)] / #[revision(indexed_seq)] / #[revision(indexed_set)] instead — the type-level map = "indexed" and seq = "indexed" forms are rejected at parse time with a diagnostic pointing at the per-field variant.

OFFSET_TABLE_MIN_LEN = 8 is the minimum entry count that triggers the prologue; below it the encoder falls back to a sequential body and the walker falls back to a linear scan.

Validation

Indexed compounds validate their prologue eagerly on walker construction:

• Offsets are strictly monotonic
• Every offset is in-range for the payload
• Indexed-map keys are strictly ascending (byte compare)

Corrupt payloads surface as typed Error variants — InvalidOptimisedTag, OptimisedOffsetOutOfRange { offset, payload_len }, OptimisedOffsetsNonMonotonic, OptimisedKeyRegionNotAscending, OptimisedSubReaderOverrun — never as panics. Error is #[non_exhaustive] so future variants do not break exhaustive matches.

Runtime requirement: BorrowedReader

The indexed walkers (IndexedStructWalker, IndexedMapWalker, IndexedSeqWalker) borrow from a &[u8] payload. To carve that payload out of a streaming Read source they require BorrowedReader. &[u8] and SliceReader implement it; pure streaming readers (file, socket) fall through to a materialised path that allocates.

Backward compatibility

Worked example: migrating a struct from legacy to optimised

A type that started life as a single legacy revision and is now being opted into the optimised encoding for new writes:

// Before — single legacy revision:
#[revisioned(revision = 1)]
struct Profile {
    id: u32,
    handle: String,
    bio: String,
}

// After — two revisions, the new one uses optimised:
#[revisioned(
    revision(1),                                        // existing on-disk data
    revision(2, optimised, indexed_struct),
)]
struct Profile {
    id: u32,
    handle: String,
    bio: String,
}

What changes:

• Existing rev-1 bytes on disk continue to decode through the revision(1) arm — the macro normalises both the legacy revision = 1 form and the explicit revision(1) form to the same internal legacy entry, so no on-disk migration is needed.
• All new writes serialise at rev 2: u16 2 | u32_le payload_length | [u32_le; 3] offset prologue | id | handle | bio. Reading those new bytes is automatic — the macro emits one decode arm per history entry.
• A walker constructed from any rev-1 or rev-2 byte stream exposes the same per-field methods (decode_id, decode_handle, decode_bio). Skip is O(1) on rev-2 (one u32_le read + advance) regardless of how big bio got.

Indexed-map / indexed-seq / indexed-set fields

For BTreeMap / Vec / BTreeSet-shaped fields that benefit from O(log n) key lookup or random-access metadata on the wire, opt the field into the indexed encoding via one of the three per-field attributes:

use std::collections::{BTreeMap, BTreeSet};
use revision::prelude::*;

#[revisioned(revision(1, optimised))]
struct Doc {
    id: u32,
    #[revision(indexed_map)]
    fields: BTreeMap<String, Value>,   // walker can binary-search keys
    summary: String,                    // default optimised serialisation
    #[revision(indexed_seq)]
    tags: Vec<String>,                  // offset-table seq
    #[revision(indexed_set)]
    roles: BTreeSet<String>,            // sorted-bytes set; membership via walker
}

Each per-field attribute routes through its trait:

Custom container types can implement the relevant trait to participate. Hash-based containers (HashMap, HashSet) sort entries by serialised key bytes on encode so the wire layout is binary-searchable on read.

At most one of these attributes may be set per field — the macro errors at compile time if you declare more than one.

Hand-rolled SerializeRevisioned impls can call the free helpers directly:

use revision::optimised::indexed::{serialize_indexed_map, IndexedMapWalker};

let mut bytes = Vec::new();
serialize_indexed_map(&my_map, &mut bytes).unwrap();

// Reader side: binary-search a key without allocating the map.
let w: IndexedMapWalker<String, u32> =
    IndexedMapWalker::from_payload(&bytes).unwrap();
let target = "bravo".as_bytes();
let value = w.find_value_bytes(|k| k.cmp(target))?.unwrap();

Note: the encoder sorts entries by their serialised key bytes before writing, which can differ from K-order when the key's SerializeRevisioned emits a length prefix that varies across keys (as String does). Round-trip is preserved because BTreeMap's DeserializeRevisioned re-inserts entries into K-order anyway.

Worked example: an enum under the optimised tag

Tag size class tells the codec how to read each variant's payload. Inline variants are one byte total on the wire; varlen variants carry a u32_le length so skip is O(1).

#[revisioned(revision(1, optimised))]
enum Event {
    #[revision(size = "inline")]
    Heartbeat,
    #[revision(size = "fixed(16)")]
    Uuid(uuid::Uuid),                 // exactly 16 bytes on the wire
    #[revision(size = "varlen")]
    Message(String),                  // u32_le length + bytes
}

// Skim variants without materialising the payload:
let bytes = revision::to_vec(&event).unwrap();
let mut r: &[u8] = &bytes;
let walker = Event::walk_revisioned(&mut r)?;

if walker.is_heartbeat() {
    // No-op; the tag was 1 byte total.
} else if walker.is_message() {
    let text = walker.decode_message()?;   // reads u32_le len, slurps body
    // ...
}

The decode_<variant> accessor works on every walker repr (Wire, OptimisedBorrowed, ConvertedOwned) — the recommended path for surrealdb-style filters that peek the variant before deciding whether to fully decode.

Limitations (current iteration)

• Walker on optimised enums exposes discriminant(), is_<variant>(), and decode_<variant>(self) directly. The consuming into_<variant> accessor (returning a borrowed sub-walker) errors on the OptimisedBorrowed and ConvertedOwned paths — that's the Walker<'r, R> GAT lifetime trap; use <variant>_view(self) -> VariantView<'r, T> to get the variant payload bytes (borrowed from the source in the common OptimisedBorrowed case), then construct your own walker from view.as_bytes() if needed.
• The type-level map = "indexed" / seq = "indexed" attributes are rejected at parse time — they're impossible to implement soundly without specialisation (the macro can't tell BTreeMap from any other field type). Use the per-field #[revision(indexed_map)] / #[revision(indexed_seq)] / #[revision(indexed_set)] attributes instead. They work today.
• fixed(N) requires the variant body to serialise to exactly N bytes under SerializeRevisioned. Use [u8; N], Uuid, fixed- width primitives under fixed-width-encoding, etc. — varint-encoded primitives have variable length and won't match. The macro emits a debug_assert_eq! in the encode arm to catch declared-vs-actual size mismatches.

Attribute spelling convention

The optimised wire format adds several attributes; they follow two shapes depending on what they declare:

• Opt-in flags are bare keywords because they're booleans — presence means "yes", absence means "no". Currently: optimised and indexed_struct at the revision level (inside #[revisioned(revision(N, ...))]); indexed_map, indexed_seq, indexed_set, fixed, specialised at the field level (inside #[revision(...)] on a field). Mixing two indexed-* markers for one field is a compile error.
• Parameterised options use key = "value" pairs because the value carries information beyond on/off: size = "inline" | "fixed(N)" | "varlen" on optimised-enum variants picks one of three classes (with an embedded byte count for fixed); start = N, end = N, convert_fn = "...", default_fn = "...", fields_name = "..." likewise take a parameter.

This split mirrors how Rust's own #[cfg(...)] works: cfg(test) is a flag, cfg(target_os = "linux") is a configuration value.