Portrait Service: AI-Driven PB-Scale Data Mining for Cost Optimization and Stability Enhancement
Talkat KubeCon + CloudNativeCon China 2025 · on Tuesday, 2025-06-10· at 17:00· Level 19 | Crystal Court I·
summary
Kuaishou describes an AI-driven “portrait service” that mines petabyte-scale Kubernetes telemetry (10TB/day from 200K+ machines and 10M+ Pods) to improve cost and stability. It uses AI to score machine health for Kubernetes scheduling (evicting or avoiding unhealthy nodes) and to build application profiles combining microarchitecture data to pinpoint compute, cache, and I/O bottlenecks; results include far fewer pod creation delays and higher IPC with much lower LLC miss rates. The roadmap includes adding AI agents to automate anomaly detection and cut manual ops by ~80%.
AI-written summary · about
related talks
35mCase Study: AI-as-a-Service on Kubernetes at Scale and In Production
31mFailing to Autoscale?
27mHarnessing Kubernetes for Scalable AI/ML Workloads
40mPanel: Kubernetes and GPU Trends for AI in Financial Services
40mOptimizing AI/ML Workloads
27mKubernetes + AI Joining Forces in the Battle Against Cancer
26mHow Ancestry Got Kubernetes to Run 2x Better Per Dollar Using AI
38mBeyond Kubernetes: Adapting To Spe...