SIG-Autoscaling Deep Dive and Q+A
summary
Explains exactly which checks block node deletion and which flags change behavior: 50% utilization threshold as a default heuristic, local-storage prevents deletion and nodes with non-evictable pods or min-node-size set stay. Shows required tooling like metrics-server, and which autoscaler flags matter (scale-down-delay, unneeded-time, parallel-deletions) plus drain timeout tuning. Includes examples of misconfiguration that keeps nodes alive, how much faster scale-down becomes after lowering delays, and how to safely enable bulk deletes.
AI-written summary · about
chapters (9)
- 0:00Autoscaling and Sky Down intro
- 2:06Sky Down cluster operation
- 5:04Utilization metrics and requirements
- 8:03Configuring thresholds
- 10:36Balancing stability and efficiency
- 13:03Optimizing Sky Down
- 15:40Testing Sky Down flags
- 17:55Advanced optimization
- 19:00Summary and Q&A
related talks
37mDeep Dive: Autoscaling SIG
29mIntro + DeepDive: SIG Scalability
01h 05mIntro + DeepDive: SIG Scalability
34mDeep Dive: SIG Autoscaling
39mSIG Autoscaling Deep Dive
37mDeep Dive into Autoscaling
30mManaging Data at Scale: Best Practices and Evolution of SIG-Apps
17mScaling Made Simple: Implementing Autoscaling for Optimal K8s Performance