RainbowBird
88625a8d84
## Summary - Add the `config_kv` schema and Drizzle migration `0020`. - Keep the ConfigKV schema, cache store, and invalidation contract in the Resource API. - Read ConfigKV through a five-minute Redis cache with PostgreSQL fallback. - Reload Router and TTS voice state through `configkv:invalidate`. - Keep Auth rate limits fixed at 20 requests per 60 seconds. ## Stack - Depends on #2294 for the Redis test implementation. - This PR adds ConfigKV-specific cache-aside and Pub/Sub tests on top of that implementation. ## Deployment Run migration `0020` before this runtime reaches production traffic. Then freeze ConfigKV writes. Audit and backfill the data with [proj-airi/backend#2](https://github.com/proj-airi/backend/pull/2). Merge [proj-airi/backend#4](https://github.com/proj-airi/backend/pull/4) first, so the fixed Auth rate-limit keys are skipped. Keep writes frozen until the hashes match and two API instances pass the Pub/Sub reload check. This PR does not run production DDL or data migration. ## Verification - `pnpm exec vitest run <ConfigKV cache store, sync subscriber, and Auth rate-limit tests>` (12 tests passed) - `pnpm -F @proj-airi/api-server typecheck` - `git diff --check` See #2294 for its frozen-install, ESLint, and 73-test verification. ## Visual changes None. This PR changes backend persistence and rate-limit wiring only. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **New Features** - Added centralized configuration storage with validation, caching, refresh, and automatic synchronization across services. - Configuration updates now refresh related language-model and text-to-speech settings automatically. - **Bug Fixes** - Improved recovery after service reconnects by clearing stale configuration and reloading current values. - Invalid or unavailable configuration data now produces clearer service-unavailable responses. - **Changes** - Authentication rate limiting now uses a consistent limit of 20 requests per minute per client. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Signed-off-by: RainbowBird <git@luoling.moe> Signed-off-by: RainbowBird <rbxin2003@outlook.com>
AIRI Backend
Project AIRI's hosted backend source lives under this folder. Workspace package
names stay stable; the directory groups service source and database ownership
while production deployment configuration remains in proj-airi/airi-railway.
Layout
apps/api: resource API, business domains, database migrations, and API runtime.apps/auth: standalone Better Auth and OIDC service.packages/auth-shared: Auth-owned database schema and principal contracts.packages/drizzle-migration: bundled migration history consumed by the API migration owner.dev/caddy: local-only public edge routing for the shared Auth/API origin.docker-compose.yaml: complete local API + Auth + PostgreSQL + Redis + Caddy stack.
Run locally
From the repository root:
pnpm dev:backend
The command uses server/docker-compose.yaml and exposes only Caddy at
http://localhost:6112.
Railway deployment
API and Auth are separate long-running Railway services built from the same repository. Keep each service's Root Directory at the repository root: both Dockerfiles copy workspace manifests and shared packages from that build context. In each Railway service, configure the Config File Path explicitly:
| Service | Config File Path | Public role | Private dependency |
|---|---|---|---|
| Resource API | /server/apps/api/railway.toml |
Product and resource API | Auth issuer and JWKS |
| Auth | /server/apps/auth/railway.toml |
Better Auth and OIDC issuer | Resource API deletion endpoint |
Each config pins its own Dockerfile, start command, /readyz healthcheck, and
watch patterns. A change only deploys a service when it changes that service,
one of its copied shared packages, or a copied root build input.
Service-to-service contract
Share database, Redis, and observability variables using Railway reference variables rather than copying secret values between services. Configure the two directional private links as follows:
| Consumer | Variable | Value source | Purpose |
|---|---|---|---|
| Resource API | AUTH_SERVER_URL |
Auth's canonical public issuer URL | JWT issuer, audience, and public JWKS identity |
| Resource API | AUTH_SERVER_INTERNAL_URL |
Auth's Railway private domain | Private JWKS fetch; it does not change issuer validation |
| Auth | PUBLIC_URL |
Auth's canonical public issuer URL | Better Auth and OIDC issuer URL; must equal API AUTH_SERVER_URL |
| Auth | RESOURCE_SERVER_URL |
API's Railway private domain | Private call before deleting a user's business data |
Set RATE_LIMIT_TRUSTED_PROXY=railway only for services directly receiving
Railway proxy traffic. Keep /internal/* private: Auth calls the API over its
private domain, and public routing must not expose the API's internal Auth
routes.
The API remains the shared database migration owner. Do not add a Railway
pre-deploy migration command to Auth, and do not make Auth startup run shared
migrations. After either service deploys, Railway must receive 200 from that
service's /readyz; deployment success alone is not sufficient evidence that
the service can reach its required dependencies.
Not included
Frontend applications remain under apps/. Cross-runtime server SDK and
protocol packages remain under packages/ because Web, Electron, plugins,
and independent services consume them.
Production Caddy routing, OpenTelemetry Collector configuration, observability
storage, and Grafana dashboards live in proj-airi/airi-railway so deployment
topology is not duplicated in the application repository.