Event-driven microservices platform — 4 FastAPI services + React frontend coordinating through RabbitMQ. One docker compose up boots the whole system: gateway, auth (JWT), orders, billing consumer, Postgres (3 DBs), RabbitMQ, frontend.

The queue under live load — publisher and consumer in balance, zero backlog:

sequenceDiagram
    participant C as Client
    participant G as API Gateway :9000
    participant A as Auth :9003
    participant O as Orders :9001
    participant R as RabbitMQ
    participant B as Billing (consumer)

    C->>A: POST /auth/token
    A-->>C: JWT
    C->>O: POST /api/orders (Bearer JWT)
    O->>O: INSERT order (Postgres)
    O->>R: publish order_created
    O-->>C: 201 {order_id}
    R->>B: order_created
    B->>B: INSERT payment (Postgres)
    B->>R: publish payment_processed

Services never call each other over HTTP — all cross-service communication is events through a durable topic exchange. Kill billing, create orders, restart billing: it catches up from the queue. That's the point of the architecture.

cp .env.example .env
docker compose up --build      # 8 containers, ~2 min first build

# 1. get a token
TOKEN=$(curl -s -X POST localhost:9003/auth/token \
  -d 'username=demo&password=Demo12345!' | jq -r .access_token)

# 2. create an order
curl -X POST localhost:9001/api/orders \
  -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{"user_id": 1, "amount": 49.90}'
# → {"order_id":6,"message":"Order created and event published"}

# 3. watch billing consume it
docker logs asyncflow_billing --tail 2
# → Payment processed: order_id=6

Aggregate health across the fleet — the gateway polls every service; the frontend renders it:

Queue consumers still need /health. Billing has no web framework — it serves liveness from a 20-line stdlib ThreadingHTTPServer so the gateway and k8s probes treat it like any other service. No FastAPI dependency for one endpoint.

Case-insensitive settings. All services read config via pydantic-settings; case_sensitive = True was silently discarding DB_HOST from compose and falling back to defaults — the kind of bug that "works on my machine" hides. Removed everywhere.

Serialization at the boundary. Events are pydantic models; model_dump_json() — not json.dumps(model_dump()) — because Decimal and datetime must survive the trip through the broker.

See docs/development.md for make commands and deployment notes.

MIT