NEW YORK — What happened Database benchmarking group OpenQuery released its late-2026 scalability review testing PostgreSQL extensions like pgvector against dedicated vector databases. While PostgreSQL handled datasets under 10 million vector embeddings with negligible latency, query throughput dropped sharply by over 70 percent once datasets crossed the 100-million mark under concurrent user load.
Why it matters As machine learning features like semantic search and contextual documentation retrieval became standard software features, development teams opted to store high-dimensional math representations (vectors) alongside traditional tables in PostgreSQL. Doing so prevents teams from needing to manage a second database engine. However, when database indexes grow larger than available system memory, performance characteristics shift abruptly, causing unexpected timeouts in production software.
Deep dive Vector search algorithms typically rely on graph-based structures such as Hierarchical Navigable Small World (HNSW) to find nearest-neighbor results quickly. Unlike conventional B-tree database indexes, which navigate small chunks of data stored neatly on fast solid-state drives, HNSW graphs require random memory access patterns across the entire graph structure. When the index size exceeds the server's RAM, the operating system is forced to swap data back and forth from disk storage. In the 2026 benchmarks, this memory thrashing caused average query latencies to jump from 14 milliseconds to over 480 milliseconds during active writes, effectively stalling synchronous web applications.
Report check The degradation cliff at the 100-million vector threshold was replicated across standard cloud virtual machines with 64GB to 128GB of RAM. What is disputed is whether newer productized quantization strategies—which compress vector dimensions to keep indexes small—compromise search accuracy too much for mission-critical enterprise discovery tasks.
Open questions Will software optimizations in disk-native vector algorithms bypass the need for massive RAM pools, or will enterprise teams with tens of millions of embeddings be pushed permanently toward specialized vector appliances?