Unlocking Kubernetes Observability: Secure, Tenant-Cen...
summary
Adobe describes a Prometheus-based multi-tenant proxy (deployed as a single CRD) that exposes curated GPU and namespace metrics (DCGM, kube-state-metrics, CPU/memory, ingress) to tenant Prometheus instances, with Tenant Exporter selection via a ConfigMap/MetricAccess CRD and quota-driven rate limits enforced by an Nginx proxy using token-bucket sliding windows. Integration uses prom-label-proxy and kube-rbac-proxy for auth, demo shows live scraping into tenant Grafana, and Adobe reports savings up to 50–60% on idle GPU cost at ~450 clusters / 150k nodes.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:07Multi-tenancy in Kubernetes
- 3:56Soft vs hard monitoring
- 5:23Prometheus federation limits
- 7:30Multi-tenant observability importance
- 8:56Adobe Kubernetes at scale
- 10:04Multi-tenant Prometheus proxy
- 13:55Cortex vs observability
- 16:43GPU optimization costs
- 21:16Demo setup overview
- 24:43Deploying monitoring stack
- 26:53Enabling proxy with CRD
- 28:40GPU metrics in Grafana
- 29:42Proxy benefits
- 30:48Q&A
related talks
43mBuilding a Multi-tenant Monitoring System with Grafana, Prometheus and Cortex
24mBanking Observability at Scale
37mMonitoring Kubernetes with Prometheus
32mDesign Observability Solution for Multi-Tenancy Environment
19mKeynote: Monitoring Kubernetes Clusters with Prometheus
46mAdvanced Kubernetes: Lesson Learned From Building..
24mPractical Kubernetes Monitoring with Prometheus
32mDemocratizing Application Monitoring with Prometheus