Repository navigation
msgpack@2: name-keyed structs #16
Description
Deferred follow-up from the compatibility-rules discussion (tracking issue in core).
msgpack@1 (today's MsgPack, rmp_serde default config) encodes a struct as a positional array, and generated Rust has no #[serde(default)]. So:
- extra trailing elements fail decoding (
LengthMismatch); - missing ones fail too, even for
Optionfields; - adding, removing or reordering any field breaks peers.
JSON tolerates all of these.
Proposal: a msgpack@2 that encodes structs as name-keyed maps (rmp_serde struct-map mode), decodes a missing optional field as absent, and ignores unknown fields. One rule set then holds across formats. The trade is size: field names on the wire.
Needs: the runtime format, conformance vectors for a new format-msgpack@2 feature, generated #[serde(default)] on optionals in comline-rust, and a TS MessagePack codec, which doesn't exist yet.
Activity
Kinflou commentedon Oct 19, 2026
A possible shape for this, from core#127: key structs by field id rather than field name. Explicit field ids would be assigned on build and written into the source, as for function ids. An id-keyed MessagePack would tolerate added and removed fields like JSON, while staying nearly as compact as today's positional
msgpack@1, as Cap'n Proto and FlatBuffers do.Kinflou commentedon Oct 20, 2026
Decided 2026-10-20 (recorded in ComlineProject/docs#64):
msgpack@2takes a parameterkeys = "name" | "id"(defaultname).name: a map keyed by field name; renaming a field is major.id: a map keyed by field id, nearly as compact as positional; renames are free; every field needs@id(the field-id follow-up of core#127). Adding or removing an optional field is minor under both; each value sets its own profile facts.