How and Why Netflix Built a Real-Time Distributed Graph: Part 3 — Querying the graph with gRPC…

This post details the query serving layer for the Real-Time Distributed Graph (RDG), focusing on handling diverse query patterns (shallow-wide vs. deep-narrow) with sub-100ms latency. Key design choices include a breadth-first traversal strategy to manage fan-out and memory trade-offs, an async-first execution pipeline to maximize I/O efficiency, selective caching of stable data (achieving 70-80% hit rates), opt-in enrichments from external services with fail-open behavior, and a default to eventual consistency. The architecture comprises a Graph Query Service, Storage Abstraction Layer (integrating KVDAL and EVCache), and an Enrichment Layer.