Leverage ML to automatically tune K8s apps
summary
Shows an ML optimizer that tunes nine parameters (four Kubernetes CPU/memory requests and limits, five JVM flags) by applying configs via the K8s API, running JMeter workloads, and scraping metrics with Elastic APM. experiment 34 finds a best config after 19 hours that cuts cost to half the baseline while keeping SLOs; experiment 14 yields ~60% cost reduction after ~8 hours and reduces autoscale latency spikes. VPA recommendations alone fail; the ML search builds real configs that work in production.
AI-written summary · about
chapters (12)
- 0:00ML-driven K8s tuning intro
- 2:12Modern app perf challenges
- 4:22K8s resource management
- 7:14CPU throttling impact
- 9:21K8s auto-scaler limits
- 11:07ML for performance tuning
- 14:16ML optimization in practice
- 17:08Case study baseline config
- 19:13Baseline perf under load
- 21:01ML config results
- 23:09High-reliability findings
- 25:38Key takeaways
related talks
22mAutomating Performance Tuning with Machine Learning
37mGetting the Optimal Service Efficiency That Autoscalers Won’t Give You
18mLet the machines optimize the machines
14mPerformance Optimization in High-Volume Transaction Systems
32mOptimizing Application Performance on Kubernetes
42mOptimize Kubernetes costs while matching your application SLO using AI
31mFailing to Autoscale?
26mDescubra las tendencias de optimización de costos en Kubernetes