Optimizing Kubernetes for Reliability and Cost-Efficiency

Introduction

• Google report into how Elite performers optimize their Kubernetes clusters

Golden Signals

• Workload rightsizing (requests vs actuals)
  • 3rd party tools e.g. Kubecost often use requests to calculate cost apportion for chargeback
• Demand-based down-scaling
  • Ensure apps can handle—startup/shutdown speed
• Cluster bin packing (requests vs allocatable)
• Discount coverage (percentage of cluster covered by e.g. spot or sustained use discounts)
  • Elite performers heavily use discounts—they understand what is going on within their clusters
  • Don’t purchase commit before rightsizing—could over-commit, need data

General Best Practices

• As a minimum: set resource requests
  • Set for workloads that actually need the reliability that burstable/guaranteed QOS brings
  • Some workloads are fine with best-effort
  • Controls many critical aspects:
    • Scheduling
    • Bin packing
    • HorizontalPodAutoscaler
    • Cluster Autoscaling
    • Cost allocation
    • Reliability
    • Performance
• Use labels and PodDisruptionBudgets
• With multi-zonal clusters—look out for inter-zone egress cost
  • Can add up—be aware of cluster/application deployment topology
  • Istio can help
• With multi-regional failover clusters
  • Ensure failover region can warm up before scheduled failover

GKE Best Practices

• Use GKE Cost Optimization tab
• Dashboards available at organizational level
• If using Autopilot—default min requests set by Autopilot if none set
  • Need to check before deployment—could under/over provision with default

References

• State of Kubernetes Cost Optimization