Skip to content

RFP: replicas, declared bidirectional replication over the transmit family #50

Description

@cardmagic

Summary

Give bidirectional replication a first-class surface: two runtimes declare a replica pair, each stages transmits for the operations the other mirrors, and each ingests the other's envelopes. The transmit family (#47, #48, shipped in #49 with the engine route) provides the mechanism; this proposal names the missing declaration. This is the feature that earns the "replica" name.

Why this is deferred

Wiring register_transmit at both ends of a pair already works today, but only for one direction at a time. Point both directions at each other and the pair loops: an ingested operation applies through an internal message, the actor's operation stages a transmit unconditionally, and the mirrored apply transmits straight back. Echo suppression is the design problem, and it deserves its own review rather than a rider on #49.

Design questions

  1. Echo suppression. A plausible default: Actor#transmit stages nothing when the current turn applies a message whose idempotency key carries the transmit: prefix, so a replayed operation never re-transmits. A multi-hop topology (A → B → C) then needs an explicit forwarding override, because suppression at B would stop the chain.
  2. Declaration API. What form earns the name: a class-level replicate_to on the actor, or a runtime-level pairing that wires register_transmit, the engine route policy, and the resolver in one declaration?
  3. Identity. A replica pair needs a bijective actor type mapping. The current resolve_actor_type / transmission_actor_type_resolver are one-way; a pair declaration should derive both directions from one mapping.
  4. Conflict posture. Per-actor ordered replay gives last-writer-wins in arrival order. Document single-writer-per-actor as the supported convention first; divergence detection (periodic state digest exchange) is a separate follow-up.
  5. Wire contract. Loop prevention across more than two runtimes likely needs an origin field in the envelope. That is a contract revision, so it lands in compatibility/transmit-envelopes.json and both runtimes together.

Plan

  1. Prototype the echo suppression rule behind the existing transmit tests, and prove a two-runtime pair converges instead of looping.
  2. Settle the declaration API against a real pair (the Rails QA app and a Node runtime already exercise both directions one way at a time).
  3. Revise the fixture file if origin lands, with consuming tests in both repositories.
  4. Documentation, roadmap, parity ledger, both changelogs, both suites green.

References

#47, #48, the shipped transmit family and engine route in #49, and solid-objects-js#17 / solid-objects-js#18.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions