Scaling Made Simple: Implementing Autoscaling for Optimal K8s Performance
summary
Explains concrete autoscaling behavior: Cluster Autoscaler reacts to unschedulable pods (example: 4‑core/8GB pod on an 8‑core/16GB node, then a 12GB request forces a new node), default deprovisioning threshold is ~50%, and HPA and KEDA scale pods by request-based CPU/memory or event/custom metrics (example: 1GB request with 70% memory trigger → add replicas at 700MB). Warns VPA unreliable for high seasonality, shows a dashboard comparing utilization, requests, and allocation, and notes sometimes fixed replicas (e.g., five) are simpler and cheaper.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:53Day Two Operations
- 1:49Autoscaling
- 2:24Scaling tools
- 3:07My expectations
- 3:46Scheduling
- 5:25Cluster Autoscaling
- 6:10HPA
- 9:26Pod Requests
- 10:39Limits
- 12:35Pet vs Cattle
- 13:39Prioritization
- 14:31Dashboard
related talks
34mIntroduction to Autoscaling
35mSIG Autoscaling Intro
29mKubernetes Autoscaling: Beyond One Size Fits All
30mAutoscaling in Kubernetes
35mAutoscaling Kubernetes Deployments: A (Mostly) Practical Guide
21mMusic Matched to Different Ways to Scale in K8s, What Could Go Wrong?
31mFailing to Autoscale?
37mDeep Dive into Autoscaling