Agent Skills: Serialization

Serialization

Overview

Converting in-memory data to bytes for storage/transport and back. The hard problems are not encode/decode calls — they are schema evolution (old and new code exchanging data safely), format choice (self-describing vs schema'd, text vs binary), and correctness traps (int64 precision, NaN, non-deterministic output, unknown-field handling). This skill is mechanism-level; assume you can call the library.

Format Selection

| Format | Schema | Self-describing | Cross-lang | Notes / when to reach for it | | --- | --- | --- | --- | --- | | JSON | none | yes | yes | Debuggable, ubiquitous. No int64/binary/date types; slow; large. Default for public HTTP APIs. | | Protobuf | IDL (.proto) | no (tag numbers only) | excellent | Compact, tag-based evolution, gRPC. Needs .proto to read bytes. | | Avro | required (writer+reader) | schema travels or via registry | good | No per-field tags/names in payload → smallest tagged binary; schema-resolution evolution. Kafka/Hadoop. | | MessagePack | none | yes | yes | "binary JSON": same data model, ~2x smaller/faster. Dynamic data without a schema. | | CBOR (RFC 8949) | none | yes | yes | IETF-standard MessagePack cousin; has a canonical form + tags. COSE/WebAuthn/DTLS. | | FlatBuffers / Cap'n Proto | IDL | no | good | Zero-copy: read fields without parsing; mmap-able. Games, low-latency IPC. | | bincode / postcard | Rust type layout | no | no (Rust↔Rust) | Fastest+smallest for Rust-only. No evolution — field order/type is the contract. | | TOML / YAML | none | yes | yes | Config, human-authored. Not for hot-path data (YAML esp. slow + footguns). |

Decision heuristics: public API / must be human-readable → JSON; internal RPC, polyglot, evolving → protobuf (gRPC); event stream with a registry → Avro; Rust-to-Rust cache/IPC, no evolution → bincode/postcard; need to read one field of a huge blob → FlatBuffers; cryptographic canonicalization → CBOR canonical or JCS.

Schema Evolution & Compatibility

The discipline that lets producers and consumers deploy independently. Precise definitions:

• Backward compatible = new reader can read data written by old writer. (You upgraded the consumer.)
• Forward compatible = old reader can read data written by new writer. (You upgraded the producer first; consumers lag.) Requires readers to ignore unknown fields.
• Full = both. Aim for full on anything with independent deploy cadence (events, shared queues, mobile clients you can't force-update).

| Change | Backward | Forward | Notes | | --- | --- | --- | --- | | Add optional field w/ default | ✅ | ✅ | The safe change. Old reader ignores it; new reader defaults it. | | Remove optional field | ✅ | ✅ | Reserve its tag/name (protobuf). | | Add enum value | ✅ | ⚠️ | Old reader must tolerate unknown → open enums (proto3) or a catch-all case. | | Widen int (int32→int64) | ✅ | ⚠️ | Forward-breaks if new values exceed old type's range. | | Rename field (JSON) | ❌ | ❌ | JSON keys by name; add-new + dual-write + drop-old instead. | | Change field type / tag number | ❌ | ❌ | Wire garbage or silent misparse. New field number instead. | | optional→required | ❌ | — | Never in proto3 (required doesn't exist — by design). |

Expand/contract (parallel change) is the only safe rename/retype: (1) add the new field, (2) write both old+new, (3) migrate readers to new, (4) stop writing old, (5) remove + reserve old. Same pattern as DB migrations.

Protocol Buffers

Field numbers are the permanent, load-bearing identity

The .proto field name is cosmetic; the tag number is what's on the wire. Rules that bite hard if violated:

• Never reuse a field number. A recycled number makes new readers misinterpret old bytes as the new field — silent corruption, no error. When deleting, reserved the number and the old name so nobody re-adds them.
• Numbers 1–15 use a 1-byte tag; 16–2047 use 2 bytes. Put hot/repeated fields in 1–15.
• Compatible type swaps (same wire type) are safe: int32/int64/uint32/uint64/bool/enum are all varint and interchangeable with sign caveats (negative int32 is 10 bytes; range overflow silently truncates). sint32/64 (zigzag) and fixed32/64 are different wire types — not interchangeable with int*. string↔bytes compatible when bytes are valid UTF-8.

syntax = "proto3";
package users.v1;                 // version in the package, not field names

message User {
  string id = 1;                  // 1–15: 1-byte tags for hot fields
  string email = 2;
  optional string phone = 4;      // `optional` = explicit presence (see below)
  repeated string roles = 5;      // packed by default in proto3
  map<string, string> metadata = 6;
  google.protobuf.Timestamp created_at = 7;

  reserved 3, 50 to 60;           // retired numbers — never reuse
  reserved "legacy_name";         // retired name
}

proto3 presence, enums, unknown fields

• Implicit presence (default proto3): a scalar at its zero value (0, "", false) is indistinguishable from unset and is not serialized. If you must tell "0" from "absent" (PATCH semantics, tri-state), mark the field optional (explicit presence, since protoc 3.15) or wrap in google.protobuf.Int32Value. Message-typed fields always have presence.
• Open enums: proto3 enums are open — an unknown numeric value round-trips as the raw int rather than erroring. The zero value must be *_UNSPECIFIED (safe default + forward-compat sentinel). Never renumber existing values.
• Unknown fields are preserved on round-trip in modern protobuf (proto3 dropped this in 3.0–3.4, restored in 3.5 / 2017). This is what makes a proxy that decodes→re-encodes forward-compatible. ⚠️ Protobuf JSON mapping drops unknown fields by default — a JSON gateway is not forward-compatible the way the binary form is.

Wire-format gotchas

• Serialization is not canonical: map entry order is unspecified, unknown fields are appended, and repeated SetSerializationDeterministic is per-process best-effort, not a cross-language guarantee. Do not hash/sign raw protobuf bytes expecting stability. Sign the exact received bytes, or canonicalize explicitly.
• required (proto2 only) is a permanent trap: you can never safely remove it (old readers reject messages missing it) — proto3 removed it deliberately.
• Large repeated/map have no built-in size cap; set decoder recursion/size limits to avoid decompression-bomb DoS on untrusted input.

gRPC & API contracts

• Partial update (PATCH) needs a google.protobuf.FieldMask, not zero-value sniffing — the mask names exactly which fields to touch, so clearing a value is expressible and unset ≠ "leave alone". Alternatively wrap fields in wrapper types.
• Prefer well-known types (Timestamp, Duration, Empty, Struct, Any, FieldMask) over hand-rolled equivalents — they have canonical JSON mappings and cross-language support.
• Choose streaming shape up front (unary / server-stream / client-stream / bidi); it is part of the wire contract and can't be changed compatibly. List responses should carry next_page_token, not rely on stream length.
• Version in the package (users.v1) so a v2 can coexist; never mutate a shipped message's field meanings.

JSON: pitfalls experts guard against

• int64/uint64 lose precision in JavaScript. JS numbers are IEEE-754 doubles; integers above 2^53−1 (Number.MAX_SAFE_INTEGER) round silently. This is why the protobuf→JSON mapping encodes 64-bit ints as strings. Transport large IDs/amounts as strings; parse with BigInt.
• NaN / Infinity are not valid JSON. JSON.stringify(NaN) → null (silent data loss); most strict parsers reject Infinity. Guard float fields before encoding.
• No canonical form. Key order, whitespace, number formatting (1e2 vs 100, -0, trailing zeros), and Unicode escaping all vary → re-serialized JSON ≠ byte-identical. Never compare/sign JSON by re-encoding (see canonical section).
• Duplicate keys are undefined. Spec allows them; most parsers keep last-wins, some first-wins → a smuggling vector across two services using different parsers. Reject duplicates on security boundaries.
• Missing types: no date (use ISO-8601 strings), no binary (base64), no integer/float distinction, no comments/trailing commas. Parsing 2024-12-19 back to a Date is your job.
• null vs undefined vs omitted are three distinct states. Decide per field: omit unknown-but-optional (skip_serializing_if), send null for known-absent. Emitting undefined in JS just omits the key — align both ends or PATCH semantics break.

Canonical / Deterministic Serialization

Needed whenever bytes are hashed, signed, deduplicated, or content-addressed. The rule: serialize once, then treat the bytes as opaque — never re-encode and expect equality.

• JSON: use RFC 8785 JCS (JSON Canonicalization Scheme) — lexicographic key sort, minimal number formatting, fixed Unicode escaping. For signing, prefer signing the raw payload bytes (as JWS/COSE do over the exact octet string) rather than JCS-normalizing, to avoid round-trip drift.
• CBOR: has a defined canonical/deterministic encoding (RFC 8949 §4.2) — used by COSE (cose-sign1) and WebAuthn. Reach for CBOR over JSON when you need standardized determinism.
• Protobuf: not canonical (see above). If you must, define an explicit canonical byte layout or hash a canonicalized projection, not the wire bytes.
• Common bug: verifying a signature by re-serializing the parsed object. Any field reorder, default-omission, or float reformat breaks it. Keep and verify against the original bytes.

Rust serde

Primary language of this project. Derive covers the common cases; the value is in the attributes and the enum-representation trade-offs.

use serde::{Deserialize, Serialize};

#[derive(Serialize, Deserialize)]
#[serde(rename_all = "camelCase", deny_unknown_fields)]   // JSON keys: userId, createdAt
struct Order {
    id: String,
    #[serde(default)]                                     // backward-compat: tolerate missing
    #[serde(skip_serializing_if = "Vec::is_empty")]      // omit empty on the wire
    items: Vec<Item>,
    #[serde(with = "time::serde::rfc3339")]               // module supplies ser+de
    created_at: OffsetDateTime,
    #[serde(skip)]                                        // never (de)serialized; needs Default
    cache: Cache,
}

Key attributes: rename / rename_all (camelCase,snake_case,SCREAMING_SNAKE_CASE,kebab-case); default / default = "path"; skip_serializing_if = "Option::is_none"; with/serialize_with/deserialize_with; flatten; tag/content/untagged; borrow (zero-copy &'a str); deny_unknown_fields; alias (accept an old key name on read).

Enum representations — pick deliberately, each has a trap

#[serde(tag = "type")]                      // internally tagged: {"type":"Text","content":"hi"}
#[serde(tag = "t", content = "c")]          // adjacently tagged: {"t":"Text","c":"hi"}
#[serde(untagged)]                          // by shape: "hi"  or  {"url":...}
// default (externally tagged):             {"Text":"hi"}

• untagged and internally-tagged buffer the whole input into an intermediate Content value: (1) measurably slower, (2) breaks zero-copy — #[serde(borrow)] &'a str can't borrow from a buffer, forcing owned String; (3) untagged tries variants in declaration order and reports only a generic "did not match any variant" — order variants specific→general and expect poor errors.
• Internally tagged can't represent newtype-of-primitive variants (Text(String)) — the tag needs a map/struct to live in; use struct variants (Text { content: String }) or adjacent tagging.
• flatten forces a map-based deserialize path. It silently breaks non-self-describing formats (bincode, most binary), is incompatible with deny_unknown_fields (flatten needs to see leftover keys), and disables serde's field-count fast path. Prefer explicit fields when the format is binary.
• deny_unknown_fields trades forward-compat for strictness — safe for internal configs, wrong for evolving external payloads where you want to ignore new fields.
• Use #[serde(other)] on a unit variant as an enum catch-all (tagged enums) so new server-side variants don't hard-fail old clients.

Non-self-describing formats (bincode/postcard) can't drive internally/adjacently/untagged enums or flatten, and have no schema evolution — reordering fields or changing a type silently misreads. Use them only for same-version Rust↔Rust.

MessagePack & CBOR

• Same data model as JSON plus real integers, binary, and (via extensions) custom types. ~2× smaller and faster than JSON; still self-describing (no schema needed).
• CBOR over MessagePack when you need an IETF standard, a canonical form (signing), or tag-based extensibility (datetime tag 0/1, bignum). MessagePack when you just want compact JSON with the widest library support.
• Both preserve unknown map keys → naturally forward-compatible for additive changes.
• ⚠️ Untrusted input: set nesting-depth and length limits; a hostile 5-byte header can claim a multi-GB array/map and OOM a naive decoder.

Performance

• Cost order (Rust, rough): bincode/postcard < protobuf ≈ MessagePack/CBOR < JSON ≪ YAML. Binary saves parse time and bytes.
• Amortize setup: reuse encoders/decoders and, for protobufjs/Ajv-style libs, load/compile the schema once — recompiling per message dominates cost.
• Stream large payloads with JSON Lines ({...}\n{...}) or length-delimited protobuf instead of one giant document — bounds memory and enables backpressure. Skip malformed JSONL lines rather than failing the whole stream.
• Compress at the transport, not per-field (gzip/zstd over the response); mining bytes by hand-shortening keys hurts readability for marginal gain vs a binary format.
• Measure with representative payloads — small-object microbenchmarks mislead; allocation/GC pressure often dominates raw encode time.

Gotchas (cross-format)

• Floats are not exact. 0.1+0.2 ≠ 0.3; round-tripping money as float drifts. Serialize currency as integer minor units (cents) or a decimal string, never f64.
• Timezones: always emit UTC ISO-8601 with offset (...Z); a bare 2024-12-19T14:30:00 is ambiguous across systems.
• Endianness / integer width only matter for hand-rolled binary — every format above handles it, but a custom DataView/byteorder layout must fix both ends.
• YAML footguns: NO/no/on/off parse as booleans (Norway problem), 1.0 may become a float, unquoted large numbers overflow. Quote ambiguous scalars; use safe_load, never yaml.load on untrusted input (arbitrary object construction).
• Trusting length prefixes / recursion depth from untrusted input = decompression-bomb / stack-overflow DoS. Cap both.

Verification checklists

Schema change:

• [ ] New protobuf fields use new numbers; removed numbers+names are reserved
• [ ] Change classified against the compatibility table (backward/forward/full) for this data's deploy cadence
• [ ] Renames/retypes done via expand/contract (add-new, dual-write, migrate, drop-old), not in place
• [ ] Readers ignore unknown fields (forward-compat) OR strictness is a deliberate internal choice
• [ ] proto3 tri-state fields use optional/wrappers, not zero-value-as-absent
• [ ] Round-trip tested: old writer→new reader and new writer→old reader

Correctness:

• [ ] 64-bit ints crossing a JSON boundary are strings (no 2^53 precision loss)
• [ ] Floats guarded for NaN/Infinity; money is integer-cents or decimal-string
• [ ] Dates are UTC ISO-8601 with offset
• [ ] Signed/hashed payloads use the original bytes or a canonical form (JCS/CBOR-canonical), never a re-serialization
• [ ] Untrusted decode has depth/length caps; YAML uses safe loader