14 min read
Nine failed requests, every Tuesday at 14:10: what a rolling update actually drops
Kubernetes reports the rollout as successful, and it is — the guarantee is about pod counts, not about the requests that were in flight. Why a well-behaved process that exits promptly on SIGTERM is the worst kind, what the preStop window is really for, and the four timeouts that have to agree before 'zero downtime' means anything.
kubernetesdevopsreliabilityperformance