Netflix Turns Operational Data into a Knowledge Graph for AI-Driven Observability
A typical application regression investigation at Netflix can take around four hours, often because engineers must reconstruct connections between systems. Netflix is building an operational ontology-driven knowledge graph to make those connections explicit. An ontology defines operational entities, their properties, and how they relate; the graph stores those relationships in a form that people and machines can use. For example, the graph can connect applications with incidents, deployments, alerts, and infrastructure dependencies. Enrichers gather information from operational systems and turn it into subject-predicate-object triples. Some links are created deterministically, while LLMs and probabilistic techniques help interpret less structured information. Validation rules constrain those outputs. The graph is stored in QuipuDB, Netflix’s distributed RDF database. It supports the company’s AutoRCA system and its effort to produce ranked, evidence-backed incident explanations. Netflix wants to extend this foundation from root-cause analysis to automated remediation and, ultimately, self-healing infrastructure. Humans will retain approval authority over proposed changes and fixes.
What is an operational ontology-driven knowledge graph, and what is Netflix using it for?
A typical application regression investigation at Netflix can take around four hours, often because engineers must reconstruct connections between systems. Netflix is building an operational ontology-driven knowledge graph to make those connections explicit. An ontology defines operational entities, their properties, and how they relate; the graph stores those relationships in a form that people and machines can use.
For example, the graph can connect applications with incidents, deployments, alerts, and infrastructure dependencies. Enrichers gather information from operational systems and turn it into subject-predicate-object triples. Some links are created deterministically, while LLMs and probabilistic techniques help interpret less structured information. Validation rules constrain those outputs.
The graph is stored in QuipuDB, Netflix’s distributed RDF database. It supports the company’s AutoRCA system and its effort to produce ranked, evidence-backed incident explanations. Netflix wants to extend this foundation from root-cause analysis to automated remediation and, ultimately, self-healing infrastructure. Humans will retain approval authority over proposed changes and fixes.
How much activity does Netflix’s infrastructure handle, and how many concurrent streams did it support during the Jake Paul–Mike Tyson fight?
Netflix says its infrastructure handles more than 2 billion requests each day and upwards of 38 million real-time logging events every second. During the November 2024 Jake Paul-Mike Tyson fight, it supported approximately 65 million concurrent streams. These figures show the scale of the operational environment the company needs to monitor.
The article does not connect those activity figures to a particular investigation or say that the fight caused an incident. Instead, it describes the numbers as context for Netflix’s wider challenge: understanding a large, interconnected environment of applications and infrastructure. The company already collects metrics, events, logs, and traces, but those signals are spread across operational systems.
Netflix is developing an ontology-driven knowledge graph to make relationships among that information explicit. The graph is intended to help engineers and AI agents investigate incidents, including application regressions. Netflix plans to use it for ranked, evidence-backed explanations, with humans retaining approval authority over proposed changes and remediation. The longer-term goal is automated remediation and self-healing infrastructure.
Why can finding the cause of an application regression take around four hours, even when Netflix collects extensive operational data?
Netflix already collects metrics, events, logs, and traces across its applications and infrastructure. But useful evidence is spread among alerting, incident management, deployment tools, and engineering communication platforms. Each system holds part of the picture, rather than one consistent account of how operational components connect.
A regression might originate in a client release, an experiment, infrastructure load shedding, or a service several dependencies downstream. Engineers often have to find and link those clues manually before they can focus on the underlying data. Netflix says a typical application regression investigation can take around four hours, with much of that time spent reconstructing relationships.
Netflix identified the missing capability as a common way to describe components and their relationships. Its Observability Ontology supplies that model, and a knowledge graph makes the relationships explicit. Enrichers bring information from operational systems into the graph. Netflix aims to use it to speed root-cause analysis, while humans retain authority over proposed changes and remediation.
How does Netflix’s system connect signals such as logs, metrics, deployments, and alerts to rank possible causes of an incident?
Netflix’s system brings scattered operational information into a shared model. The Observability Ontology defines entities and their relationships, while the knowledge graph links information from applications, incidents, deployments, alerts, and infrastructure. This lets the system consider signals together rather than treating each tool’s data as an isolated clue.
The correlation engine examines 13 types of signals, including log errors, metric anomalies, traces, alerts, incidents, deployments, configuration changes, experiments, and code changes. Enrichers convert source information into structured relationships. The engine searches for possible causes and evaluates evidence such as temporal proximity and severity, then produces ranked hypotheses.
Some enrichment is deterministic; other stages use LLMs and probabilistic techniques to interpret less structured information. Netflix constrains those outputs with validation rules, including entity types, URI naming conventions, and checks against its software catalogue. The system is intended to produce evidence-backed explanations, not automatically approve fixes: humans retain approval authority over proposed changes and remediation.
What could this system change about how engineers investigate incidents and apply fixes?
Today, engineers can spend around four hours investigating a typical application regression, much of it reconstructing relationships across systems. Netflix’s knowledge graph makes operational connections explicit and available to both engineers and machines. Its purpose is to help investigations focus sooner on likely causes rather than on manually gathering scattered context.
The correlation engine examines 13 signal types and ranks potential causal explanations using evidence such as timing and severity. Netflix also captures incident discussions, including hypotheses, rejected explanations, and mitigation decisions. That human reasoning can remain in the operational knowledge base and inform future detection and analysis instead of disappearing when a conversation ends.
Netflix’s internal AutoRCA system uses the graph for root-cause analysis. The company intends to produce ranked, evidence-backed explanations and preserve human approval authority over proposed changes and remediation. It ultimately wants to extend the capability into automated remediation and self-healing infrastructure. Netflix also plans to explore links with its Unified Data Architecture, connecting operational information to broader business and technology domains.
How does a shared ontology help AI agents analyze different parts of an incident and contribute to the same investigation?
Netflix’s ontology gives AI agents a common model of operational entities and relationships. That shared vocabulary matters because agents may analyze different parts of an incident, such as traces, logs, or deployments. Without a common interface, their findings could be difficult to connect or compare.
With the ontology, each agent can contribute information about the same operational entities without needing to understand another agent’s internal implementation. The knowledge graph stores those connections alongside information about applications, incidents, alerts, and infrastructure dependencies. Netflix contrasts this with specialist agents exchanging loosely structured findings.
The graph also supports AutoRCA, Netflix’s internal agentic AI system for root-cause analysis. Netflix aims to use it to generate ranked, evidence-backed incident explanations. The shared model is a foundation for that work, not a claim that agents already remediate incidents independently. Humans retain approval authority over proposed changes and remediation, while automated remediation and self-healing remain longer-term goals.
How do knowledge graphs use subject-predicate-object triples to represent relationships between real-world things?
A subject-predicate-object triple represents a relationship as three parts: the subject is one thing, the predicate names the relationship, and the object is the thing it relates to. This structure gives a knowledge graph a direct way to store links between entities, rather than keeping each fact as an unconnected piece of text.
For example, an operational graph could represent an application as the subject, “depends on” as the predicate, and a downstream service as the object. That example illustrates the format; Netflix’s article describes its graph as representing operational relationships among applications, incidents, deployments, alerts, and infrastructure dependencies. Enrichers convert source information into these triples.
Netflix uses deterministic enrichment for some information and LLMs or probabilistic techniques for less structured material. Validation rules constrain the resulting data. The graph stores the structured relationships in QuipuDB and makes them available to engineers and AI agents. The article identifies RDF among the standards Netflix uses for its ontology and graph.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web