microservices-design（微服务设计）

Microservices Design

Service Boundaries

Define services around business capabilities, not technical layers. Each service owns its data store and exposes a clear API contract.

order-service/       -> owns orders table, publishes OrderCreated events
inventory-service/   -> owns inventory table, subscribes to OrderCreated
payment-service/     -> owns payments table, handles payment processing
notification-service -> stateless, subscribes to events, sends emails/SMS

Event-Driven Communication

interface DomainEvent {
  eventId: string;
  eventType: string;
  aggregateId: string;
  timestamp: string;
  version: number;
  payload: Record<string, unknown>;
}

const orderCreatedEvent: DomainEvent = {
  eventId: crypto.randomUUID(),
  eventType: "order.created",
  aggregateId: orderId,
  timestamp: new Date().toISOString(),
  version: 1,
  payload: { customerId, items, totalAmount },
};

await broker.publish("orders", orderCreatedEvent);

async function handleOrderCreated(event: DomainEvent) {
  const { items } = event.payload as OrderPayload;

  for (const item of items) {
    await db.inventory.update({
      where: { productId: item.productId },
      data: { quantity: { decrement: item.quantity } },
    });
  }

  await markEventProcessed(event.eventId);
}

Use idempotency keys (eventId) to handle duplicate deliveries safely.

Saga Pattern (Orchestration)

class OrderSaga {
  private steps: SagaStep[] = [
    {
      name: "reserveInventory",
      execute: (ctx) => inventoryService.reserve(ctx.items),
      compensate: (ctx) => inventoryService.release(ctx.items),
    },
    {
      name: "processPayment",
      execute: (ctx) => paymentService.charge(ctx.customerId, ctx.amount),
      compensate: (ctx) => paymentService.refund(ctx.paymentId),
    },
    {
      name: "confirmOrder",
      execute: (ctx) => orderService.confirm(ctx.orderId),
      compensate: (ctx) => orderService.cancel(ctx.orderId),
    },
  ];

  async run(context: SagaContext): Promise<void> {
    const completed: SagaStep[] = [];

    for (const step of this.steps) {
      try {
        const result = await step.execute(context);
        Object.assign(context, result);
        completed.push(step);
      } catch (error) {
        for (const s of completed.reverse()) {
          await s.compensate(context);
        }
        throw new SagaFailedError(step.name, error);
      }
    }
  }
}

API Gateway Pattern

# Kong or similar gateway config
services:
  - name: orders
    url: http://order-service:3000
    routes:
      - paths: ["/api/v1/orders"]
        methods: [GET, POST]
    plugins:
      - name: rate-limiting
        config:
          minute: 100
      - name: jwt
      - name: correlation-id

  - name: users
    url: http://user-service:3000
    routes:
      - paths: ["/api/v1/users"]
    plugins:
      - name: rate-limiting
        config:
          minute: 200

Health Check Pattern

app.get("/health", async (req, res) => {
  const checks = {
    database: await checkDatabase(),
    cache: await checkRedis(),
    broker: await checkMessageBroker(),
  };

  const healthy = Object.values(checks).every(c => c.status === "up");

  res.status(healthy ? 200 : 503).json({
    status: healthy ? "healthy" : "degraded",
    checks,
    version: process.env.APP_VERSION,
    uptime: process.uptime(),
  });
});

Anti-Patterns

• Sharing a database between services (tight coupling)
• Synchronous HTTP chains across multiple services (cascading failures)
• Building a distributed monolith (services cannot deploy independently)
• Missing circuit breakers on inter-service calls
• Not implementing idempotency for event handlers
• Using distributed transactions instead of sagas

Checklist

• Each service owns its own data store
• Services communicate via events for async workflows
• Saga pattern used for multi-service transactions with compensation
• Circuit breakers protect against cascading failures
• API gateway handles routing, rate limiting, and authentication
• Health check endpoints report dependency status
• Event handlers are idempotent (safe to process duplicates)
• Services can be deployed and scaled independently