gke-compute-classes

What problem does it solve?

Configuring GKE ComputeClasses manually is error-prone, with common pitfalls including scheduling conflicts, node stockouts, misaligned cost optimizations, and unenforced access controls that lead to wasted spend or workload downtime. This Skill eliminates those risks by providing validated guidance, guardrails, and pre-built templates for end-to-end ComputeClass management.

Core Features & Use Cases

• Cost Optimization: Configure Spot VMs with on-demand fallback, align machine families with existing Committed Use Discounts (CUDs) and reservations, and implement priority ladders to minimize spend while maintaining availability.
• Workload Targeting: Set up ComputeClasses for GPU/TPU workloads, specific machine families (c3, c4, n4), and zonal resource colocation for performance-sensitive applications.
• Troubleshooting: Diagnose and fix pending pods, scheduling failures, and ComputeClass misconfigurations, including common issues like missing GPU tolerations, reservation affinity errors, and stockout cascades.
• Governance: Restrict ComputeClass access using RBAC for CRUD operations and ValidatingAdmissionPolicy for consumption control, preventing unauthorized teams from modifying or requesting high-cost classes.
• Use Case Example: For a team running latency-sensitive AI inference on GKE, use this Skill to build a ComputeClass that prioritizes GPU reservations, falls back to on-demand nodes, and enforces guardrails to avoid Spot preemption for production traffic.

Quick Start

Use the gke-compute-classes skill to generate a validated ComputeClass YAML template for your GPU inference workload that aligns with your existing committed use discounts and includes proper tolerations for GPU nodes.