Proactive cluster autoscaling in Kubernetes
summary
Shows a concrete pattern: create a placeholder pod with low priority that fills a node so the cluster autoscaler requests a replacement node as soon as real pods need space, and the placeholder is evicted on demand. Gives timings and tools: HPA reaction ~90s, autoscaler request ~30s plus ~3–4 minutes to boot and ~30s to join, the placeholder demo cuts cold‑start to about reduced cold-start from ~3–4 minutes to ~7 seconds. Also ships an instance calculator, a scheduler simulator library and a Black Friday game to test behavior; the live demo mostly works but sometimes glitches.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 2:00Kubernetes scaling myths
- 3:43Kubernetes scheduler
- 6:07Node sizing and reservations
- 8:14Cluster autoscaling delays
- 9:17Black Friday scaling game
- 12:53Proactive scaling strategies
- 13:56Instance calculator
- 15:00Placeholder pods
- 23:10How placeholder pods work
- 24:32Proactive scaling in practice
- 28:02Cost and efficiency
- 28:44Workload prioritization
related talks
28mCluster autoscaling in Kubernetes
29mProactive cluster autoscaling in Kubernetes
34mIntroduction to Autoscaling
34mPredictive autoscaling patterns in Kubernetes
35mAutoscaling Kubernetes Deployments: A (Mostly) Practical Guide
17mScaling Made Simple: Implementing Autoscaling for Optimal K8s Performance
35mSIG Autoscaling Intro
31mFailing to Autoscale?