Skip to content

submitqueue: port internal queue contracts to proto (submitqueue/core/messagequeue) #356

Closed

Description

@behinddwalls

Problem

The repo's queue-contract convention (doc/rfc/messagequeue-contract.md) is proto3 payloads serialized as protojson, with a contract package owning both the messages and their topic_keys bindings — api/runway/messagequeue/ is the reference example and stovepipe/core/messagequeue/ follows it. The submitqueue domain has not been ported: there is no submitqueue/core/messagequeue/ package, and 11 of its 12 topics still carry hand-rolled JSON via ToBytes/FromBytes methods on entity types.

Current state per topic:

Topic key Payload Serialization
start RequestID entity JSON
cancel CancelRequest entity JSON
validate RequestID entity JSON
batch RequestID entity JSON
score BatchID entity JSON
speculate BatchID entity JSON
prioritize QueueID entity JSON
build BatchID entity JSON
buildsignal BuildID entity JSON
conclude BatchID entity JSON
log RequestLog entity JSON
merge MergeRequest ✅ protojson (api/runway/messagequeue — correctly borrowed; runway owns that queue's contract)

Consequences of the split: no additive-evolution guarantees (protojson's unknown-field discard, UPPER_SNAKE enums, int64-as-string are only enforced on merge), no topic_keys binding or contract test for internal topics, and queue-payload serialization concerns living on domain entities (which are supposed to be pure data).

Work

  1. Create submitqueue/core/messagequeue/ mirroring the reference layout: proto/ sources, committed protopb/, and the generic protojson glue (Marshal / Unmarshal[T] / TopicKeys) — same as stovepipe/core/messagequeue/.
  2. Define proto messages for each payload shape (request-ID carrier, batch-ID carrier, build-ID carrier, queue-ID carrier, cancel request, request log) and bind each topic key to its message via the topic_keys option from api/base/messagequeue.
  3. Add the contract test: protojson round-trip per message + every topic key bound to exactly one message.
  4. Migrate producers and consumers topic by topic (gateway Land/Cancel publishes, all orchestrator stage controllers, core/request.PublishLog, and the DLQ controllers/republish tooling — anything touching msg.Payload).
  5. Once no queue payload uses them, remove ToBytes/FromBytes from the submitqueue entity types (RequestID, BatchID, BuildID, QueueID, CancelRequest, RequestLog, and the full-entity variants) so entities go back to being pure data.
  6. Bazel visibility stays domain-scoped per the internal-contract rule.

Caveats

  • Wire-format flip: protojson output (snake_case fields, int64-as-string) differs from the current encoding/json entity output, so in-flight messages from before a cutover won't parse after it. Port topic-by-topic with drained queues (fine at the current pre-prod stage), or give consumers a transitional dual-format read if needed.
  • The log topic crosses gateway ↔ orchestrator but both are submitqueue-domain services, so the internal contract location (submitqueue/core/messagequeue/) is correct — it must not go under api/.
  • The message-ID convention is a separate concern tracked in messagequeue: entity-ID message IDs + publish dedup silently drop wake-up events; use intent-scoped message IDs #352 (intent-scoped message IDs); porting payloads to proto neither depends on nor resolves it, but call sites will be touched by both — coordinate to avoid churn.

Related

Activity

  1. behinddwalls commented on Jul 14, 2026

    @behinddwalls
    CollaboratorAuthor

    Linear: CODEM-305

  2. behinddwalls commented on Oct 5, 2026

    @behinddwalls
    CollaboratorAuthor

    The bulk migration is already landed in #703: internal proto payloads and generated bindings, protojson helpers, topic-key bindings/contract tests, producer/consumer and DLQ migration, and domain-scoped Bazel visibility.

    The final remaining cleanup is in #780: retire the unused full-entity Request/Batch/Build JSON serializers and their serializer-only tests, add unknown-field compatibility coverage for all 11 current internal payloads, and document the original cutover/dead-letter caveats. No new wire-format or message-ID change is introduced.

    Validation on the cleanup branch:

    • All 55 SubmitQueue/domain-service test targets pass.
    • Native lint, module tidy, and Gazelle checks pass.
    • Fake-provider E2E and isolated real-Git single-change/stack cases pass.
    • The full real-Git suite hits a local Docker-mounted repository ownership failure, also reproduced on unchanged main at 276bfcc2; the PR documents this and the transient local-test/tooling limitations.

    #780 contains Closes #356, and GitHub confirms the closing-issue association. It is currently a draft and requires review before merge, so this issue remains open until the final cleanup lands.

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions