KEDA, HPA and VPA? An Introduction to Scaling
summary
A Node.js Fastify demo app that blocks 100 ms per request runs on three replicas and shows CPU ≈1 core/pod, event loop lag spikes up to 3s, and many timeouts under Gatling load. Adding the under-pressure plugin returns 503s to protect pods; HPA scales from 3 to about 5–6 replicas and reduces latency. Moving CPU work to worker threads lowers event-loop utilization from ~75% to ~20% and cuts max/99th latencies by ~70%. KEDA (via Prometheus on OpenShift) is configured to scale on event-loop lag >30 ms for smarter autoscaling.
AI-written summary · about
chapters (8)
- 0:00Introduction to scaling
- 1:05Speaker introduction
- 8:49Non-blocking IO challenges
- 17:37Handling back pressure
- 21:32Horizontal Pod Autoscaler basics
- 29:18Worker threads in Node.js
- 38:01KEDA and custom metrics
- 43:50Q&A session
related talks
35mKEDA - Real Time And Serverless Scaling In Kubernetes
37mAuto-scaling party:VPA, HPA, KEDA, nodes, how do they dance?-Nic Vermande & Oluebube Princess Egbuna
08mGetting Creative with HPA Scaling in Kubernetes
37mKEDA: Unlocking Advanced Event-Driven Scaling for Kubernetes
37mKubernetes Event-driven Autoscaling with KEDA
46mScaling EDA Workloads with Kubernetes, KEDA & Karpenter
49mEvent-driven autoscaling for Serverless Java
32mScaling Kubernetes Application to 1M Users