Evolution of Aegis: Fault Diagnosis for AI Model Training Service in Production
summary
Aegis is a production fault diagnosis system for AI model training that evolves from enhanced legacy tools to a customized collective communication library (CCL) for runtime localization without changing customer code. It adds a pre-delivery validation step and a degradation detector using 20+ metrics and z‑score outlier analysis. Deployed for a year at Alibaba Cloud, Aegis claims >97% reduction in idle diagnosis time, 84% fewer restarts and 71% less performance degradation.
AI-written summary · about
chapters (9)
- 0:00Aegis introduction
- 1:42Large-scale AI training challenges
- 3:51Existing diagnosis limits
- 5:34Aegis Phase one
- 8:01Procedure-aware runtime diagnosis
- 10:30Performance degradation system
- 12:38Pre-delivery checks
- 13:44Aegis production results
- 15:31Conclusion and future
related talks
25mMonitoring AI with AI
08mProduction Evals For Agentic AI Systems
17mAI-enhanced Systems Engineering for the Development of complex Medical Devices
26mManufacturing Alerts To ReActive Agents: Self-Evolving AI Agents on K8S for Predictive…
15mAI Gardening: Harvesting and Pruning the AI Tree
43mAI-Driven Quality Engineering Practices
38mAI Just Happened to Me : Help!!
28mImplementing AI-Driven Observability in Healthcare
