Vector Search Beyond the Hype: Real-World Implementation and Performance Insights

Vector Search Beyond the Hype: Real-World Implementation and Performance Insights - Samuel Herman, Datastax Vector search has emerged as a cornerstone technology in modern search applications, with numerous indexing technologies promising scalability and performance. However, the practical implementation of these solutions reveals important nuances that aren’t captured in marketing materials. This session provides a data-driven exploration of vector search implementation challenges through the lens of developing and contributing the jVector library as a KNN engine plugin for OpenSearch. We’ll begin by establishing quantitative benchmarks to guide decisions on when and how to leverage vector indices, followed by an examination of critical operational design principles including code portability, logging, instrumentation, and hardware optimization patterns. The presentation will demystify advanced vector search concepts such as product quantization, fused asynchronous distance computation, and DiskANN implementations. Most valuably, we’ll share real-world performance profiling insights, revealing which optimizations delivered meaningful improvements versus theoretical benefits. Attendees will gain practical knowledge for evaluating vector search effectively in a production environments.