Kubernetes probes: startup, readiness, and liveness

Operational decision

A search API loads an index before accepting traffic. Its startup endpoint fails until initialization finishes; readiness checks whether it can serve the current request class; liveness checks for a stuck process that a restart could plausibly repair. Do not make liveness depend on a shared database outage, because that can restart every replica during the same downstream failure. The fragment shows probe configuration for a container exposing port 8080. The handlers and timing need load tests: a probe that times out only during normal peak CPU can turn a healthy service into a restart loop. Watch startup duration and restart counts after each release, and verify readiness goes false during a deliberate shutdown before the process exits.

startupProbe:
  httpGet: {path: /health/startup, port: 8080}
  periodSeconds: 3
  failureThreshold: 40
readinessProbe:
  httpGet: {path: /health/ready, port: 8080}
  periodSeconds: 5
livenessProbe:
  httpGet: {path: /health/live, port: 8080}
  periodSeconds: 10

Cost and verification

The startup settings allow roughly two minutes for initialization before failure, aside from request timeouts and scheduling effects. Frequent probes create network and handler work; keep endpoints cheap and bounded. Readiness failure removes the Pod from normal service routing but does not restart it, while liveness failure can trigger a restart. A restart may make a transient dependency outage worse. Test thresholds under realistic CPU pressure and during termination, not only on a quiet laptop.

Common Mistakes

• Do not use the same dependency-heavy check for liveness and readiness.
• Do not set startup thresholds below ordinary cold-start time.
• Do not assume readiness catches incorrect business responses.

Connected lessons

• DevOps Tutorial
• Kubernetes Deployment: rolling update capacity
• Observability: join metrics, logs, and traces
• Incident response: contain impact, then learn

Operational follow-up

• Kubernetes Service discovery: selectors, endpoints, and DNS

Advanced follow-up

• Load-balancer health: a green socket is not a working transaction

Advanced follow-up

• Required configuration keys: fail startup before accepting traffic