Health Checks: Startup, Readiness, Liveness

1let startupComplete = false
2let shuttingDown = false
3 
4// Refreshed on a timer, NOT per probe — probes must not load the dependency.
5const depStatus = { db: 'unknown' as 'ok' | 'degraded' | 'unknown', checkedAt: 0 }
6setInterval(async () => {
7  try { await pool.query('SELECT 1'); depStatus.db = 'ok' }
8  catch { depStatus.db = 'degraded' }
9  depStatus.checkedAt = Date.now()
10}, 5_000).unref()
11 
12// LIVENESS — no I/O, no dependencies. Only things a restart can fix.
13app.get('/livez', (_req, res) => {
14  const lagMs = eventLoopLag()          // runtime-specific wedge signal
15  res.status(lagMs < 5_000 ? 200 : 503).end()
16})
17 
18// READINESS — should this instance receive traffic right now?
19app.get('/readyz', (_req, res) => {
20  if (!startupComplete || shuttingDown) return res.status(503).end()
21  // Only gate on the DB if nothing this service serves works without it.
22  if (depStatus.db === 'degraded' && !SERVES_CACHED_READS) return res.status(503).end()
23  res.status(200).end()
24})
25 
26// STARTUP — one-time initialisation; a generous failureThreshold is correct.
27app.get('/startupz', (_req, res) => res.status(startupComplete ? 200 : 503).end())
28 
29// Draining: readiness goes false FIRST, then in-flight work completes.
30process.on('SIGTERM', async () => {
31  shuttingDown = true                   // /readyz now 503; LB drains us
32  await sleep(DRAIN_GRACE_MS)           // let the LB notice before we stop accepting
33  await server.close()
34  await logger.flush()
35  process.exit(0)
36})

The DRAIN_GRACE_MS sleep is the part most implementations omit. Between readiness turning false and the load balancer removing the instance there is a window measured in probe periods, and requests arriving in that window are dropped if the server closes immediately.