Demystifying Kubernetes Resource Management
summary
Live demos use 1 worker node with 2 CPUs and 2GiB and a Google resource-consumer to run three pods: no-req, req-only, req+limit while watching Grafana and kubectl top. Claims: requests convert to CFS CPU shares and the scheduler treats pods without requests as 200m/200Mi, so a pod that requests 500m can starve best-effort pods; CPU is “compressed” (burstable) while memory is not and triggers OOM kills or kube evictions; LimitRanges and ResourceQuotas enforce per-namespace minima/maxima; StormForge is proposed to auto-tune at scale.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 4:21Why resource management matters
- 8:17Requests and limits in Kubernetes
- 12:18Live demo: CPU contention
- 16:05CPU limits role
- 17:52Scheduler uses requests
- 20:23Memory management and QoS
- 24:39Live demo: memory scenarios
- 30:30Multi-tenant clusters with limits
- 32:16Setting requests at scale
- 33:02Closing and Q&A
related talks
29mKubernetes Workload Resiliency in Action: Beyond Basics
32mKubernetes Namespaces Unleashed: Unlocking the Full Potential of You...
41mPanel Discussion: State of Art for Enabling Performance Sensitive Workloads and What We Need Next
32mOptimizing Resource Usage in Kubernetes
44mA Practical Guide to Container Scheduling
38mThe Elements of Kubernetes - Foundational Concepts for Apps Running on Kubernetes
31mSIG Scheduling Update: Transition From Pod To Workload Scheduling
06mLightning Talk: OpenKruise: Advancing Kubernetes Workload Auto...