High-Quality Telemetry Data: The Key to Effective AI Root Cause Analysis

What does it actually take for AI to perform effective root cause analysis in distributed systems? In this session, Nikolay Sivko, Co-Founder and CEO of Coroot, breaks down the critical role telemetry data plays in making AI-powered RCA work in practice.

The talk starts with an honest look at what large language models can and cannot do. While LLMs bring general knowledge of how systems behave, they know nothing about your specific application unless you provide the right context. That means the quality of your telemetry data is the deciding factor between useful AI insights and unhelpful noise.

Nikolay walks through what comprehensive telemetry actually looks like for AI root cause analysis, covering service topology and service maps, SLIs across every service and database, logs, Kubernetes and cloud events, resource metrics, network metrics, runtime metrics, database-specific metrics, profiles, and more. He also covers the different approaches to gathering that data, including manual instrumentation, OpenTelemetry SDK-based instrumentation, and eBPF-based instrumentation.

A key insight from the session is that feeding raw metrics, logs, and traces directly to an LLM performs poorly. The more effective approach is preprocessing the data and providing LLMs only with findings and relevant context, which is also significantly more cost-effective in terms of token usage.

The session closes with real benchmark data on what AI-assisted RCA can reliably detect today, including infrastructure, orchestration, network, database, and code change-related failures, and where the gaps still remain.

If you're an SRE or DevOps engineer exploring AI root cause analysis for your observability stack, this talk offers a grounded, practical framework for what works and why.