AIOps, are we there yet?
summary
Claims that simple statistics plus cohort-based baselines and clustering of multiple metrics catch variance and precursors to outages at scale, requiring roughly thousands of recurring daily workloads for stable signals. A datacenter cooling ML model reports a 40% PUE improvement in one deployment. Workflows for automated remediation remain experimental, code is proprietary but papers and talks exist, and observability-focused ML that surfaces actionable correlations proves more practical and less risky than agentic control.
AI-written summary · about
chapters (14)
- 0:00Introduction and background
- 2:04SRE for ML and ML apps
- 3:51AI for SRE and AIOps
- 5:12Challenges in distributed systems
- 8:00Limitations of SLIs and SLOs
- 11:48Statistical workload characterization
- 14:37ML for data center efficiency
- 16:22Why AIOps matters now
- 17:26Acknowledgements and resources
- 18:09Q&A: Smaller-scale applicability
- 19:12Q&A: Workload clustering methods
- 20:38Q&A: Open source and techniques
- 21:20Q&A: Agentic workflows examples
- 22:27Q&A: Promising AIOps techniques
related talks
34mWhat If the Promise of AIOps Was True?
17mOptimizing Multi-Cloud Platforms with AI Ops and FinOps
27mAI-Powered Reliability: Stop Fighting Fires, Prevent Them
02mAIOPS – How can machine learning help in IT operations
29mIs AIOps the Future of Operations? Real Use Cases From the Trenches
08mSelf-Healing Infrastructure and AI-Driven Code Evolution
30mThe Robots Are Coming (But Not Yet), or, Don't Make Smart People Do Dumb Things
17mBuilding More Resilient Cloud-Native Platforms