…ation The fleet domain layer (registry, infra registry, config store, discovery publisher) existed but had no importers outside its own package: no routes, no subscriptions, no config keys and no startup wiring. Auxiliary cc-* services had no way to register, heartbeat, discover peers or pull their configuration. This exposes it. REST (machine-to-machine, mounted under /api with RoleAPI): POST /api/fleet/register/cluster/ 201 {instanceId, configRevision} POST /api/fleet/register/infra/ 201 {instanceId, configRevision} POST /api/fleet/heartbeat/{id} 204, 404 for an unknown instance GET /api/fleet/config/{id} 200 + ETag, 304, or 204 DELETE /api/fleet/deregister/{id} 204, idempotent The config pull serves the merged blob with its content-hash revision as an ETag, so a client sending If-None-Match gets a header-only round trip in the steady state. ErrNoConfig answers 204 rather than 404 because no authored configuration is a normal state that every member polls, and the revision is acknowledged only when it differs from the stored one, which keeps polling from generating one write per member per interval. NATS carries heartbeats only. Registration, deregistration and configuration stay on authenticated REST, so no publisher on the unauthenticated broker can create, resurrect or terminate a service identity; any other function tag is rejected and counted. The message timestamp is ignored in favour of the server clock so a skewed or hostile publisher cannot park last_heartbeat in the future and evade the stale sweep. Heartbeats are coalesced by a single flusher into at most one transaction per second, mirroring BatchUpdateNodeStates, because every individual update is an fsync on a single-writer database. Discovery rosters are published on registration change (debounced) in addition to the periodic tick, so a new member is discoverable in milliseconds. The subsystem shuts down before the NATS connection is closed so the publisher cannot tick against a closed client. Configuration lives in a new main.fleet block rather than main.api-subjects, whose required subject keys would otherwise force job and node subscriptions on an operator who only wants fleet. An absent block disables routes, goroutines and subscription together. Earlier commits on this branch that build the fleet domain layer this API exposes: 5f94ae3 Initial template for registration service ecb1ba2 feat(fleet): add service type enum and infra scope eaaa7c2 feat(fleet): add hierarchical config store 98c0c17 feat(fleet): add NATS discovery roster publisher Closes #568 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Add a discovery and deployment service #568

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions