Netflix Builds Observability Ontology and RDF Knowledge Graph to Cut 4-Hour Troubleshooting

Netflix built an Observability Ontology and RDF knowledge graph stored in QuipuDB, using type constraints to convert non-deterministic LLM and fuzzy-matching outputs into deterministic, SPARQL-queryable facts that feed AutoRCA. The system addresses roughly four-hour investigations in which engineers rebuild entity mappings that individual tools never expose as discoverable relationships, spanning a scale of over 2 billion requests per day and about 38 million real-time log events per second.
Event version 1.
Key insights
Ontology as shared vocabulary
The ontology defines types, properties and allowed relations as subject-predicate-object triples, such as api-gateway ops:ownedBy "Team Bedrock". A single query can answer which incidents relate to Bedrock applications in us-east-1.
Rejecting a generic Service type
Netflix unified entities as ops:Application validated against a software catalog of about 49,000 entries, and unvalidated strings do not become applications. Bidirectional dependency edges are capped to prevent popular services from overwhelming the graph.
Entropy ladder from non-determinism
LLM summaries, fuzzy matching and probabilistic causal chains are unstable, so the ontology narrows output space through enumerations into deterministic triples. A five-layer pipeline progressively reduces the share of information still trapped in unstructured form.
Correlation engine as evidence-backed hypothesis
The correlation engine reads 13 signal classes within roughly a 10-minute window and finds paths by temporal proximity and severity, discarding correlation chains with confidence below 0.6. Chains and supporting graph records are saved with scores such as 0.87.
AutoRCA keeps humans in control
Multiple specialist agents collaborate around one problem using the ontology as common vocabulary, while alerts still fire, humans still decide and code is still written by people. Each hypothesis carries evidence because hypotheses can still be wrong.