Operating within Normal Parameters: Monitoring Kubernetes
summary
Defines SLOs like 99% control-plane uptime and pod start p99 <5s, then builds a minimal in‑cluster monitoring stack using Prometheus, kube-state-metrics, node_exporter and cAdvisor. Prometheus survives an API‑server outage in production, so the stack works for alerting. Practical rules: detect an offline node by Prometheus up==0, flag slow kubelets with scrape duration >2s, find quota issues with namespace CPU >95% or overprovisioning if actual CPU <35%. Notes API‑server histogram bucketing capped at 8s (fixed in Kubernetes 1.14).
AI-written summary · about
chapters (11)
- 0:00Intro
- 3:31Defining Service Level Objectives
- 9:08What sources of metrics are available?
- 13:19What types of metrics are available?
- 14:40A minimal monitoring stack for Kubernetes
- 16:14Run your monitoring stack on Kubernetes!
- 18:17Let's not worry about high availability!
- 21:03Metric analysis, aggregation, visualization
- 22:32Service Degradation: Node is down
- 24:23Service Degradation: Customer can't launch Pods
- 26:08Service Degradation: API Server is slow
related talks
25mAn SLO-Driven Approach to Enhance Kubernetes Cluster Reliability
19mKeynote: Monitoring Kubernetes Clusters with Prometheus
37mWeighing a Cloud: Measuring Your Kubernetes Clusters
24mPractical Kubernetes Monitoring with Prometheus
37mMonitoring Kubernetes with Prometheus
23mDon't Let Users Find Your Outages: Synthetic Monitoring for Kubernetes Platforms
37mUsing SLOs for Continuous Performance Optimizations of Your K8s Workloads
39m15 KUBERNETES FAILURE POINTS YOU SHOULD WATCH