Netflix built a knowledge graph to help AI agents find production bugs

A Netflix on-call engineer takes about four hours to find out why a production feature started failing. Most of that time is not reading data. It is working out how the systems involved relate to each other. Netflix wants its AI agents to answer that question in one place, so engineers can skip the rebuild.

InfoQ reports on 13 November that Netflix built an ontology-driven knowledge graph for this. The report is by Craig Risi. It is based on a post on the Netflix Tech Blog called "Ontology-driven observability: building an end-to-end knowledge graph at Netflix scale."

What the graph holds

The graph describes operational entities and how they relate. The entities include applications, incidents, deployments, alerts and infrastructure. The graph uses standard formats: RDF, OWL, SHACL and SPARQL. The data sits in QuipuDB, which the article calls "Netflix's distributed RDF database built on Rust's redb storage engine."

The scale the graph has to cover is large. Netflix handles more than 2 billion requests a day. It takes in up to 38 million real-time logging events every second. During the Jake Paul-Mike Tyson fight in November 2024, it carried about 65 million concurrent streams.

Why it is a shared surface for agents

The reason to put all of this in one shape is the AI agents. Netflix calls its system AutoRCA, for automated root-cause analysis. One agent reads logs. Another reads traces. Another reads deployments. Each writes what it finds into the same graph, in the same shape.

InfoQ quotes the Netflix post. The agents use the ontology as a common interface. Each one can write about operational entities into the graph. It does not have to read how any other agent works inside. A new agent added next month can read everything the old agents wrote.

A correlation engine then runs across what the agents wrote. It looks at 13 signal types. The named nine are: log errors, metric anomalies, traces, alerts, incidents, deployments, configuration changes, experiments and code changes. The article does not name the other four.

What the article does not say

The InfoQ piece does not say how many of the four-hour investigations AutoRCA has actually cut, or by how much. Netflix has not published a measured gain yet. The Netflix Tech Blog post describes what the system is and what it is for. It does not give a before and after number.

Netflix says next it will move from automated analysis into automated remediation and self-healing infrastructure. That is a forward goal, not a working system today.

For a team building AI agents that run in production, the shape is the lesson, not the scale. A shared graph with an agreed ontology lets a team add an agent without rewriting the others. Each agent's output belongs to the system, not to itself.

Source

Craig Risi, InfoQ, "Netflix Turns Operational Data Into a Knowledge Graph for AI-Driven Observability". Netflix Tech Blog, Ontology-driven observability: building an end-to-end knowledge graph at Netflix scale.