您如何实现优雅关闭？

Question

您如何实现优雅关闭？

Accepted Answer

**优雅关闭** 是指服务器在收到终止信号时以干净的方式停止 — 完成飞行中的请求、关闭连接 (DB 等) 并释放资源 — 而不是突然杀死它并丢弃活动工作。这对于零停机部署和可靠性至关重要。

## 为什么突然关闭是个问题

```text
Without graceful shutdown, when the process is killed (deploy, scale-down, crash):
  ✗ In-flight requests are dropped → users get errors
  ✗ DB transactions left incomplete
  ✗ Connections not closed cleanly → leaks/locks
```

在部署和自动扩展期间，进程不断被停止 — 以干净的方式进行这一操作是使部署无缝的原因。

## 实现

```js
import http from "http";
const server = http.createServer(app);
server.listen(3000);

async function shutdown(signal) {
  console.log(`${signal} received — shutting down gracefully`);

// 1. Stop accepting NEW connections; finish existing ones
  server.close(async () => {
    // 2. This callback runs once all active requests have completed
    await db.end();            // close database pool
    await redis.quit();        // close other connections
    console.log("clean shutdown complete");
    process.exit(0);           // exit successfully
  });

// 3. Safety net: force-exit if cleanup hangs too long
  setTimeout(() => {
    console.error("forced shutdown — cleanup timed out");
    process.exit(1);
  }, 10_000).unref();
}

// Listen for termination signals
process.on("SIGTERM", () => shutdown("SIGTERM")); // sent by orchestrators (k8s, Docker)
process.on("SIGINT", () => shutdown("SIGINT"));   // Ctrl+C in dev
```

顺序：捕获信号 → `server.close()` (停止新连接、排空现有连接) → 关闭 DB/Redis/其他资源 → 退出。**超时降级** 在某些内容挂起时强制退出，因此卡住的连接不会永远阻止关闭。

## 信号

```text
SIGTERM → the standard "please terminate" signal from orchestrators on deploy/scale-down
SIGINT  → Ctrl+C (development)
SIGKILL → cannot be caught (force kill) — graceful shutdown can't help here
```

## 在容器化环境中

```text
Kubernetes/Docker send SIGTERM, wait a grace period (e.g. 30s), THEN SIGKILL.
Your app must finish within that window → keep the shutdown timeout shorter.
Readiness probes should mark the pod unready so the load balancer stops sending traffic.
```

## 为什么这很重要

优雅关闭是在现代（尤其是容器化/编排）环境中实现 **零停机部署** 和可靠操作的关键，在这些环境中进程会定期停止和重新启动。

没有它，每次部署或缩减都会丢弃活动请求并冒着数据不一致的风险。

实现 catch-signal → stop-accepting → drain → close-resources → exit 序列 (带有超时安全网) 是生产就绪的 Node 服务的标志，也是一个经常被忽视但很重要的运维问题。