Cluster Autoscaler Evolution
summary
Cluster Autoscaler outlines concrete changes: DRA and CSI node fixes land and are DRA production ready (v1.35), a unit test framework and benchmark arrive, and two new APIs — capacity buffer (beta) and capacity quota (alpha) — are implemented to trade latency for cost and to cap provisioned resources. A repo split extracts core autoscaling logic from cloud provider integrations, a proposal called Work at the Worst Scheduler replaces CA’s scheduler simulation with scheduler‑autoscaler APIs, and GKE/AKS forks’ useful features are being upstreamed.
AI-written summary · about
chapters (17)
- 0:00Cluster Autoscaler intro
- 1:05What Cluster Autoscaler does
- 3:37HPA and VPA integration
- 4:39Cluster Autoscaler history
- 7:27Recent milestones
- 9:37Carpenter project impact
- 11:01Cloud provider forks challenges
- 13:51Maintainability issues
- 17:03Scheduling features solutions
- 19:51Dynamic Resource Allocation
- 22:19New APIs for capacity
- 25:10Node info visibility
- 27:35Device utilization testing
- 29:20Workload-Aware Scheduler
- 33:18Refactoring for maintainability
- 34:40GKE features upstreaming
- 35:21Q&A
related talks
39mCluster API & Cluster Autoscaler
34mDeep Dive: SIG Autoscaling
39mSIG Autoscaling Deep Dive
35mAutoscaling Kubernetes Deployments: A (Mostly) Practical Guide
29mKubernetes Autoscaling: Beyond One Size Fits All
34mIntroduction to Autoscaling
37mDeep Dive into Autoscaling
24mSIG-Autoscaling: Introduction + HPA/VPA Updates