Netflix Deploys Federated Graph Search
Netflix is using a Federated Graph Search architecture to unify search and recommendations across multiple, independently managed data domains. This approach enhances scalability by avoiding a single monolithic data store while allowing for unified ranking. The company highlighted its system architecture in a recent video, underscoring that its robust A/B testing platform is a fundamental pillar for product innovation and personalization.
- The move to a federated architecture was driven by the explosive growth of Netflix Studio, which by 2019 had become a major production house, causing its monolithic API to become a development bottleneck. This new architecture, called Studio Edge, now supports 50-60 internal applications through a massive graph with 150 subgraphs, over 3000 types, and more than 2800 queries and mutations. - To make the federated graph searchable, Netflix's Content Engineering team created Studio Search, a platform that indexes portions of the graph in Elasticsearch. This allows teams to query for entities based on attributes of related entities, even when the data is owned and served by different microservices. - A key feature built on top of the search platform is "reverse search," which finds saved user queries that match a changed or updated document. This enables use cases like automatically notifying a Post Production Coordinator when a movie they are tracking meets a specific new criteria, without constantly re-querying all saved searches. - Netflix is now integrating Large Language Models (LLMs) to allow for natural language search across its federated enterprise data, moving beyond its structured Domain Specific Language (DSL). This initiative uses a Retrieval-Augmented Generation (RAG) pattern to convert ambiguous user questions into structured queries. - The federated system is an evolution from Netflix's earlier API strategies, which included a flexible but complex layer of numerous lightweight APIs and a graph language called Falcor, before attempting a single graph monolith. The current approach uses Apollo Federation to allow for distributed ownership of the graph via subgraphs. - In its broader recommendation systems, Netflix is developing a foundation model inspired by the success of LLMs in NLP. This aims to centralize member preference learning from comprehensive interaction histories, allowing innovations to be transferred more easily across different recommendation models. - All changes, from UI redesigns to recommendation algorithms and even the artwork used for titles, are subjected to rigorous A/B testing. This data-driven approach ensures that product decisions are guided by user behavior, with tests sometimes revealing that a new thumbnail can increase viewing of a title by 20-30%. - To ensure software updates don't degrade the user experience, Netflix employs a form of A/B testing called "canary tests." These tests deploy new software versions to a small subset of users to monitor key metrics, like the time delay between hitting play and video start, before a full rollout.
