Cost-effective Compute Clusters with Spot and Pre-emptible Instances
summary
Kubernetes clusters mix GKE preemptible node pools and kops with AWS spot pools (maxPrice in instance groups) to push 60–80% lower hourly costs for transient capacity. Use node selectors, node affinity/anti-affinity, taints/tolerations, Horizontal Pod Autoscaler plus cluster autoscaling backed by ASGs so burst traffic lands on spot nodes and pods reschedule when nodes die. Demo shows 5–6 minute node provisioning and cleanup, GCE preemptibles die ≈24h, and a 12-day two-node test costs $16 vs $24 (projected ~$13,000/yr saving at scale).
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:45About Platform Systems
- 1:53Introducing spot & preemptible instances
- 6:39Bidding strategies
- 7:58Amazon EC2 Spot Instances
- 11:13Implementation Mechanisms - GKE
- 13:41Implementation Mechanisms-kops with AWS
- 22:01Application scheduling considerations
- 23:27Supporting K8s mechanisms
- 31:40Case Study / Experiment
related talks
21mLessons Learned Running GKE Clusters on Spot Instances
26mScaling K8s Nodes Without Breaking the Bank or Your Sanity
33mUncover Kubernetes cost-optimization trends
34mKubernetes: Reducing Infrastructure cost and complexity
33mThe Big Kubernetes Mistake Paying More Than Necessary
27mScaling K8s Clusters Without Losing Your Mind or Money
34mReduce Costs and Optimize on Your Cloud with Karpenter
40mHow an AI driven approach reduces cloud cost and makes your Kubernetes infrastructure autonomous