gke-upgrades

Provide your cluster mode (Standard or Autopilot), current and target versions, release channel, and workload types. The Skill produces a plan covering version compatibility, sequential upgrade paths, node pool strategy, workload readiness, and rollback procedures.

Surge upgrades are the default, using maxSurge and maxUnavailable settings per pool. Blue-green suits mission-critical workloads needing fast rollback, while autoscaled blue-green (Preview) avoids pre-provisioning double capacity.

Common causes are PDBs with zero allowed disruptions, insufficient capacity for pending pods, bare pods blocking drains, admission webhooks rejecting pods, and PVC attachment failures. It may also be paused by a closed maintenance window, which is a safe mixed-version state.

No-upgrades exclusions are capped at 90 days each, with at most 3 per cluster and 48 hours of maintenance availability required per rolling 92-day window. For longer freezes, use no_minor_or_node_upgrades scope, which runs until the minor version's End of Support.

Use rolling upgrades with maxSurge=0 since GPU reservations lack spare capacity, and test driver and CUDA compatibility in staging first. If regressions occur, pin the previous driver version or roll back the node pool to the prior GKE version.

Control plane patch downgrades are supported directly, and minor rollback is only possible during step one of the 2-step minor upgrade. Other control plane minor rollbacks require GKE Support, while node pools can be downgraded directly or via a replacement pool.