gke-reliability

This skill is a reference for improving the reliability and high availability of GKE workloads and clusters. It solves the problem of workloads suffering avoidable downtime during node upgrades, autoscaler scale-downs, zone failures, and slow application startups. It provides a set of golden-path defaults (regional clusters across multiple zones, surge upgrades, auto-repair, auto-upgrade, the regular release channel, and stateful HA) and then walks through the specific Kubernetes primitives that make workloads resilient.

The skill covers verifying cluster high availability, creating Pod Disruption Budgets to guarantee minimum pod availability during voluntary disruptions, and configuring liveness, readiness, and startup probes with explicit timeouts to control traffic routing and restarts. It explains graceful shutdown using SIGTERM handling, terminationGracePeriodSeconds, and preStop sleep hooks for load balancer deregistration, and it demonstrates topology spread constraints to distribute pods across zones and nodes. A replica guidance table recommends minimums by workload type, and a best-practices section reinforces regional clusters, PDBs for every workload with two or more replicas, probes with explicit timeouts, zone spreading, and scheduling upgrades during low-traffic windows. It provides both preferred MCP tool calls (get_cluster, get_k8s_resource, describe_k8s_resource, apply_k8s_manifest, list_k8s_events) and kubectl or gcloud fallbacks.

It targets platform engineers, SREs, and application teams running production workloads on GKE. Typical use cases include hardening deployments for high availability, adding PDBs and health probes, and distributing pods across failure domains. It explicitly defers disaster recovery and full cluster backups to the gke-backup-dr skill.