Practical Capacity Planning using Prometheus
summary
Shows that Prometheus recording rules and simple stats prevent surprise bottlenecks: a Redis fleet suffers a single-core pin (one core 100% while average CPU is ~30%), so measuring service_component_saturation_ratio (max aggregator) surfaces true hotspots. Alerts fire at 90% for 5 minutes via Alertmanager, dashboards rank the worst risks in Grafana, and forecasts use a 7‑day rolling average plus predict_linear + 2σ over 2 weeks to give actionable, weeks‑ahead warnings.
AI-written summary · about
chapters (14)
- 0:00Introduction to capacity planning
- 1:11Why capacity planning matters
- 3:03Real-world resource saturation example
- 7:41Takeaways from the Redis incident
- 9:49Measuring resource saturation
- 11:57Prometheus setup for saturation metrics
- 15:52Key metrics and dashboards
- 17:46Analyzing Redis CPU saturation
- 21:33Alerting on saturation thresholds
- 22:15Forecasting saturation trends
- 25:04Improving prediction models
- 28:18Prometheus implementation of forecasting
- 30:46Prioritizing potential bottlenecks
- 32:51Wrapping up and future plans
related talks
08mPrometheus in 5 minutes
30mIntro: Prometheus
29mParanoid Observability –Success Stories and Misfortunes observing an observability stack
05mUsing Prometheus to Automate Power Capacity Planning
29m“If you Don’t Monitor your Infrastructure, you Don’t Own it!” Regain Control Thanks to Prometheus
43mFinding the Golden Signals with Prometheus
01h 19mIntro + Deep Dive: Prometheus
02h 43mPrometheus from A to Z