High-Throughput Graph Abstraction at Netflix: Part I

Netflix's Graph Abstraction is an OLTP-focused graph system handling up to 10 million operations per second across 650 TB of graph data with low latency. Built on top of existing Netflix data abstractions (Key-Value, TimeSeries, EVCache), it uses a Property Graph model with strongly typed schemas, forward/reverse edge indexes, and adjacency list storage. The architecture separates edge links from properties to prevent wide-row issues in Cassandra, employs write-aside and read-aside caching strategies, and enforces eventual consistency via Kafka-based entropy repair and async node deletions. A Gremlin-inspired gRPC traversal API supports multi-hop graph queries with filtering, sorting, and property selection pushdown. Single-hop traversals achieve single-digit millisecond p99 latency, while 2-hop traversals with high fanout stay under 50ms at p90.