Treat the control plane as a product
Upgrades, add-on lifecycles, and API deprecations are not “infra chores” — they are customer-facing changes for every team on the cluster. Publish an upgrade calendar, test against representative workloads, and measure time-to-patch when CVEs land.
Leaders should insist on one supported version skew policy and a named owner for etcd backups, API deprecation tracking, and cluster add-ons (CNI, ingress, metrics). If those live in tribal knowledge, your bus factor is the same as your on-call roster.
Observability is a contract, not a dashboard pile
Golden signals per workload, consistent label schemas, and alert routes that page the owning team — not the platform team by default. Platform provides the plumbing; service teams own SLOs and runbooks.
Review whether your alerts fire on user-visible failure modes. CPU graphs that look fine while customers time out are a leadership failure, not a tooling gap.
Pod safety and rollout discipline
Readiness probes that lie, missing PDBs, and bursty rollouts during business hours are how minor deploys become major incidents. Standardize progressive delivery patterns your teams can adopt without heroics.
Invest in game days for node loss, AZ failure, and control-plane degradation — especially after autoscaling or multi-cluster changes.