A rollout runs through cleanly, every pod is Ready, and yet requests get lost. Why deregistration runs in parallel to the shutdown, why only POST is affected, and how the preStop hook, the grace period and graceful shutdown have to fit together.
A service that returns nothing but errors looks exactly like a service with no traffic. Why that happens, how a zero series derived from the denominator fixes it, why vector(0) is not enough, and what dashboard and alert rule look like in Terraform.
A green deploy proves the service starts, nothing more. Using an order service as the example, this article walks the path to production readiness as a risk-driven process in four rounds: functional correctness, law and tax, security and operations. With the AI trap where tests only mirror the code, and arc42 as the frame.
