Every deployment produces a burst of 502s. What do you check?

Summary

Readiness probes (new pods getting traffic before they're ready), graceful shutdown (does the app receive and handle SIGTERM: exec-form CMD, draining in-flight requests), the gap between SIGTERM and load balancers updating (a preStop delay), and keep-alive timeouts between the load balancer and the app.

Longer Answer

Pod termination durations equal to the grace period mean SIGTERM was ignored (often a shell-form CMD making /bin/sh PID 1). Error spikes aligned with deploy markers confirm the pattern. Long-lived connections need explicit draining.

Key Points

• Readiness probes.
• Exec-form CMD and SIGTERM handling.
• preStop delay for endpoint propagation.
• Keep-alive alignment.

Deep dive: DevOps in Production: Every Deploy Drops Requests

What interviewers look for

Covering both the startup and shutdown sides of a rollout.