Netflix is building an operational knowledge graph that makes relationships among applications, deployments, alerts, incidents and infrastructure dependencies explicit. The goal is to reduce time spent reconstructing context before engineers can diagnose a production problem.
The platform combines an ontology—a formal description of entities and relationships—with a graph stored in Netflix’s QuipuDB. “Enricher” processes convert data from monitoring and engineering systems into structured triples. Some enrichment is deterministic, while language models interpret less structured material under validation rules for entity types, names and software-catalog references.
A correlation engine examines 13 kinds of signals, including logs, metrics, traces, configuration changes, experiments and code changes. It ranks possible causal links using evidence such as timing and severity. Incident discussions also enter the graph, preserving hypotheses, rejected explanations and mitigation decisions that would otherwise disappear in chat history.
Netflix says the graph supports AutoRCA, its internal agent system for root-cause analysis. Specialist agents can contribute findings about shared operational entities without exchanging loosely formatted prose. The company ultimately wants automated remediation and self-healing infrastructure, but those are future directions. Current plans keep human approval over proposed changes, an important boundary when probabilistic enrichment can still create an incorrect relationship.