How and Why We Rebuilt Auto-scaling in OpenFaaS with Prometheus
Talkat PrometheusDay Europe 2022 · on Tuesday, 2022-05-17· at 14:05· Pavilion 3, Room F | Level 2 | Central Forum + Live Stream, Feria Valencia, Spain + Live Stream·
summary
OpenFaaS replaces Alertmanager-driven alerts with an autoscaler daemon that queries Prometheus and kubelet/cAdvisor, supporting three concrete strategies: in-flight request mode (queue‑depth style capacity), tunable RPS, and CPU-based scaling. Recording rules collapse many PromQL joins into a single metric so the scaler computes desired replicas (example: 1,800 mCPU ÷ 500 mCPU target = 3.6 → 4 pods). Per-function scale-to-zero timeouts, dashboards, and production use (DockerCon load test, customer ML and log ingestion) prove the redesign works.
AI-written summary · about
chapters (7)
- 0:00Start
- 8:47Scale to Xero
- 11:06Hpa V2
- 11:21Edge Cases on Scale to Zero
- 13:05How Does the New Version Work
- 15:52Pod Usage Total Seconds
- 24:59Graceful Shutdown
related talks
16mAuto-scaling in Openstack without Telemetry
37mServerless Functions for Beginners
41mEvent-driven autoscaling on Kubernetes with KEDA and Azure Functions
19mDiscovering serverless with OpenFaaS
37mKEDA: Unlocking Advanced Event-Driven Scaling for Kubernetes
37mKubernetes Event-driven Autoscaling with KEDA
35mGetting Beyond FaaS: The PLONK Stack for Kubernetes Developers
39mFaaS: A Contemporary, Comparative Study