Optimizing Resource Usage in Kubernetes
summary
Adobe Experience Manager runs on more than 40 Kubernetes clusters and hundreds of thousands of environments; engineers report cluster autoscaler tuning delivers roughly 30–50% node-cost savings, switching some workloads to ARM nodes cuts ~15–20% more, and RPS-based HPA reduces active pods during the day by up to ~75%. A custom hibernation system scales dev sandboxes to zero (removing Ingress/objects) because JVMs take ~5 minutes to start. Notes: VPA in-place is alpha, etcd can hit ~8GB and default 110 pods/node is increased to ~200.
AI-written summary · about
chapters (15)
- 0:00Introduction to talk
- 1:04About our product AEM
- 3:38Managing thousands of AEM environments
- 5:24CPU and memory requests limits
- 6:29Java specific CPU challenges
- 10:12Cluster autoscaler strategies
- 13:06Vertical Pod Autoscaler issues
- 15:14Horizontal Pod Autoscaler use
- 18:07VPA vs HPA for Java workloads
- 18:30Cost savings with ARM nodes
- 19:35Hibernation of unused environments
- 22:47ETCD limits and API bottlenecks
- 25:15Max pods per node tuning
- 25:58Summary of optimizations
- 26:43Q&A session
related talks
36mHow Adobe is Optimizing Resource Usage in Kubernetes
15mDemo: Resource(Cost) optimization on planet scale Kubernetes clusters.
32mScaling Kubernetes Application to 1M Users
29mRight-Size Your Kubernetes Deployments
47mBuilding sustainable architectures with Kubernetes
17mScaling Made Simple: Implementing Autoscaling for Optimal K8s Performance
15mOptimizing Application Performance on Kubernetes
31mNew Cloud Foundry Architectures and Features and their Performance Impact