Scale Smarter Not Harder: How Extending Cluster Autoscaler Saves Millions
Talkat KubeCon + CloudNativeCon Europe 2025 · on Wednesday, 2025-04-02· at 17:45· Level 0 | ICC Capital Hall | Room J·
summary
Datadog builds an instance type adviser that runs Kubernetes scheduling simulations and benchmarks CPU/memory/network to score every candidate instance, exposes scores via CRDs/configmaps to a gRPC expander, and authors an automated node group set controller plus an instance-score controller to create fallback node groups only when needed. The system cuts one cluster’s instance count and cost by ~60%, reduces cluster-autoscaler P99 loop from minutes to <10s, and unlocks ~$4M potential savings across dozens of clusters (≈$2M realized in an early rollout).
AI-written summary · about
chapters (21)
- 0:00Scaling smarter at Datadog
- 1:04Node autoscaling explained
- 3:12Cluster Autoscaler fit
- 3:52Expanders drive instance selection
- 5:59Bin packing inefficiencies
- 7:26Built-in expanders limits
- 8:28gRPC expander for flexibility
- 10:14Instance type adviser
- 11:58Cost vs performance benchmarks
- 13:03Reliability and instance selection
- 14:28Optimal instance type tools
- 14:49Scaling with gRPC
- 16:36Validating optimization
- 17:39Automation at scale
- 19:05Instance score controller
- 20:28Node group limits and fallbacks
- 21:53Node group set controller
- 23:19Closing autoscaling loop
- 25:06Real-world impact and savings
- 26:11Next steps for platform
- 27:16Q&A
related talks
31mFailing to Autoscale?
27mScaling K8s Clusters Without Losing Your Mind or Money
26mScaling K8s Nodes Without Breaking the Bank or Your Sanity
14mCut Costs, Not Performance: Implement Kubernetes Cost Optimization Strategies
29mKubernetes Autoscaling: Beyond One Size Fits All
22mIntelligent Autoscaling in Enterprise...
17mScaling Made Simple: Implementing Autoscaling for Optimal K8s Performance
33mUncover Kubernetes cost-optimization trends