Pod lifecycle and graceful shutdown

Key points

• Phases are Pending, Running, then Succeeded or Failed. Terminating is a timestamp, not a phase.
• On delete, endpoint removal and SIGTERM start at the same time. Nothing promises the proxies update first.
• A preStop sleep of a few seconds covers that gap, which is why the pattern turns up in every production chart.
• terminationGracePeriodSeconds (30 by default) is the countdown to SIGKILL. Make it longer than your slowest request.
• Set server.shutdown=graceful in Spring Boot so SIGTERM drains in-flight requests instead of cutting them off.
• A PodDisruptionBudget stops a node drain or upgrade from removing too many replicas at once.

Example

spec:
  terminationGracePeriodSeconds: 45   # countdown to SIGKILL
  containers:
    - name: app
      image: ghcr.io/acme/orders:1.4.2
      lifecycle:
        preStop:
          exec:
            command: ["sh", "-c", "sleep 5"]   # let kube-proxy catch up
---
apiVersion: policy/v1
kind: PodDisruptionBudget
metadata:
  name: orders
spec:
  minAvailable: 2                # a drain may never go below this
  selector:
    matchLabels:
      app: orders
---
# application.yml
# server:
#   shutdown: graceful
# spring:
#   lifecycle:
#     timeout-per-shutdown-phase: 30s

SIGTERM is a request and SIGKILL is not. preStop buys the seconds the proxies need.