ODSP2135 Pre-recorded On demand
AI SRE in production: What actually gets automated safely
Arunav Sarkar
Summary
From Microsoft's AI summary
Arunav Sarkar of StackGen discusses AI in site reliability engineering and AIOps. He argues AI will not fully replace SRE teams because context is needed. He outlines a maturity model from alert triage to root cause analysis, guided remediation and restricted execution, with self-healing only in bounded domains. He reports a Kafka sandbox recovery cut from four hours to about fifteen minutes, and an etcd fix from 45 minutes to four. He also describes a failure where an unrestricted agent wiped a GPU node, and recommends policy engines, sandboxes and human sign-off. No availability or pricing was stated.
Session description
Microsoft's catalogue description
AIOps has moved past the demo stage and into real on-call rotations, but its effectiveness sits somewhere between the hype and the skepticism. This talk is a field report from running LLM-driven remediation in production: what it fixed, what it broke, and where automation has to stop and human judgment has to take over.