Talk Deep Dive
summary
Compares traditional cluster autoscaler to a policy-aware nodeless controller that chooses spot/preemptable VMs or Fargate based on app SLOs, and runs as an independent controller on the Kubernetes control plane so existing Ray clusters on Kubernetes and Helm charts deploy unchanged. Claims nodeless spins nodes up faster, removes instance-type tuning, worked out of the box on an Oracle cluster, and plans controlled experiments measuring cost savings and GPU performance across multiple cloud vendors.
AI-written summary · about
chapters (6)
- 0:00Nodeless and Ray intro
- 1:13Nodeless vs Kubernetes auto scaling
- 3:22Cost savings experiments
- 5:28Virtual Kubelet compliance
- 7:40Nodeless Kubernetes deployment
- 8:42Future roadmap
related talks
21mWhy should you bother about cluster autoscaling- Karpenter
39mCluster API & Cluster Autoscaler
28mScale Smarter Not Harder: How Extending Cluster Autoscaler Saves Millions
31mFailing to Autoscale?
06mAutoscaling GPU Clusters Anywhere — Hyperscalers, Neoclouds & Ba...
34mDeep Dive: SIG Autoscaling
24mDynamic Big Data in the Cloud with Kubernetes
27mScaling K8s Clusters Without Losing Your Mind or Money
