FinOps at scale: Lessons from scaling Kubernetes workloads
summary
Organizations scale Kubernetes and see average node CPU fall from 10% (2024) to 8% (2025) and memory from 24% to 20%, driving huge wasted EC2 costs; a 3-node estimate is $478 vs 10 nodes $1,343 (an $865 gap) using infracost. Concrete fixes: use Castai benchmarks, rightsize requests/limits with Vertical Pod Autoscaler, control Horizontal Pod Autoscaler / Karpenter / KEDA, enforce tagging via TFlint in CI, delete retained EBS volumes, and stop paying avoidable charges like $500 RDS extended support.
AI-written summary · about
chapters (18)
- 0:00Introduction and story
- 4:16Speaker introduction
- 6:27Cloud migration and complexity
- 9:17Kubernetes as cost driver
- 11:23CPU and memory utilization stats
- 13:33Overprovisioning in clusters
- 14:13Demo: Cluster cost comparison
- 18:07FinOps framework phases
- 23:04FinOps maturity levels
- 26:58Lesson: Overprovisioning fixes
- 29:22Lesson: Autoscaling strategies
- 32:30Story: Ephemeral environments issue
- 35:17Lesson: Hidden resource costs
- 37:22Lesson: Resource tagging
- 39:52Demo: Enforcing tags with TFLint
- 40:15Lesson: Shift-left FinOps
- 41:39Lesson: RDS extended support
- 43:01Final lesson: Double down on visibility
related talks
33mUncover Kubernetes cost-optimization trends
15mDemo: Resource(Cost) optimization on planet scale Kubernetes clusters.
21mLet your Kubernetes environment flourish
35mWhy Kubernetes Can't Get Around FinOps – Cost Management Best Pra...
33mThe Big Kubernetes Mistake Paying More Than Necessary
46mFinOps: Optimize for K8s
47mOptimizing Kubernetes Resource Requests/Limits for Cost-Efficiency and Latency
29mThe Frugal k8s Engineer: Strategies for Amazon EKS