Pods Under Pressure: Understadning PSI for K8s
summary
Linux PSI (available since 2018) reports sum/full/total pressure for CPU, memory and IO per process averaged over 10s/1m/5m and appears in Kubernetes 1.34 beta, 1.36 GA via the kubelet summary API or cAdvisor Prometheus endpoint. PSI flags stalled time before latency rises, so using it as a custom metric for HPA/KDA gives earlier, performance-driven autoscaling. Note: full exists only for memory; CPU shows sum only. Node autoscalers like Carpenter don’t yet consume PSI directly; combine PSI with right-sized requests rather than CPU limits.
AI-written summary · about
chapters (11)
- 0:00Introduction and pressure analogy
- 1:17Human performance under pressure
- 3:08Engineering pressure in tech industry
- 5:40Designing apps for pressure resilience
- 7:03Measuring performance metrics
- 9:12Limitations of utilization metrics
- 11:22Introducing PSI metrics
- 13:54PSI in Kubernetes and cloud costs
- 14:57Using PSI for better autoscaling
- 16:00Practical PSI implementation
- 17:52Closing and Q&A session
related talks
30mHorizontal Pod Autoscaler Reloaded - Scale on Custom Metrics
35mIntro: SIG Autoscaling
29mPower Efficiency Aware Kubernetes Scheduler
47mKEDA, HPA and VPA? An Introduction to Scaling
17mScaling Made Simple: Implementing Autoscaling for Optimal K8s Performance
47mOptimizing Kubernetes Resource Requests/Limits for Cost-Efficiency and Latency
35mDeep Dive: Kubernetes (Instrumentation) SIG
35mSIG Autoscaling Intro