Ring’s Billion-Scale Semantic Video Search with Amazon RDS for PostgreSQL and pgvector

This article details how Ring built a billion-scale semantic video search system using Amazon RDS for PostgreSQL and pgvector, enabling natural-language queries like "dog in my backyard" across billions of video embeddings.

• Ring stores 100-200 billion embeddings across 9 AWS Regions with sub-2-second query latency
• Rejected purpose-built vector databases as prohibitively expensive; S3-backed approach failed latency requirements
• Chose RDS PostgreSQL with pgvector for operational simplicity and cost efficiency
• Implemented user-based table partitioning for query performance and data isolation
• Uses brute-force parallel scans instead of vector indexes to guarantee 100% recall
• Distributes data across 4 PostgreSQL clusters per region using assignment-based routing
• EBS-optimized instances with aggressive parallelism (16 workers) unlock full bandwidth utilization
• Cold-start latency solved via pg_prewarm and read-optimized instances
• Production achieves P50 ~200ms, P95-P99 ~600ms latency with 80% read workload
• Future plans include evaluating Aurora PostgreSQL for simplified multi-cluster operations

Ring demonstrates that PostgreSQL with pgvector can handle extreme-scale vector workloads through thoughtful partitioning, parallel execution, and deep infrastructure understanding rather than specialized vector databases.