How Meta Uses Llama2 to Accelerate Incident Response and Root‑Cause Analysis in AIOps

With the rapid growth of digital transformation, traditional operations struggle to keep up, prompting the emergence of AIOps to improve efficiency, cut costs, and increase system stability. Key AIOps capabilities include intelligent monitoring, automated fault diagnosis, capacity planning, log analysis, and AI‑driven decision support.

Meta faces massive scale and complexity, serving billions of users and many enterprises. Its incident response workflow historically involved manual, chaotic hand‑offs between existing responders and new responders, leading to delays and information overload.

To streamline onboarding, Meta introduced an AI‑assisted context‑ready process using Llama2. When a new responder opens an incident, a real‑time generated summary aggregates all relevant data, and a chat assistant powered by Llama2 answers follow‑up questions, reducing the need for manual information gathering.

For root‑cause analysis, Meta categorizes incidents into changes, overload, or hardware failures. Focusing on change‑related incidents, Meta leverages a 700‑million‑parameter fine‑tuned Llama2 model that, together with historical incident data, ranks the most likely code or configuration changes. The model processes large change sets by grouping them into batches of 20, iteratively narrowing down to a handful of candidates.

Results show that the AI‑assisted workflow improves responder readiness and root‑cause discovery: about 86% of incidents benefit from the system, and within minutes of detection there is a 42% chance of identifying the underlying cause.

Meta also defines three principles for using Generative AI in operations: transparency to build trust, logical explainability for user comprehension, and ensuring actionable recommendations.