Reproduce “An In-Depth Study of Filter-Agnostic Vector Search on a PostgreSQL Database System: [Experiments and Analysis]” on OpenResearch

About this paper

Duo Lu, Helena Caminal, Manos Chatzakis, Yannis Papakonstantinou, Yannis Chronis, Vaibhav Jain, Fatma Özcan

This study analyzes filter-agnostic vector search algorithms within a PostgreSQL-compatible database, quantifying the impact of system-level overheads on performance. It demonstrates that clustering-based methods often outperform graph-based ones, especially for lower-dimensional data, and that optimal algorithm choice is heavily influenced by filter selectivity and data correlation, with system costs frequently dominating distance computations.

Key insights

• System-level overheads (e.g., page access, data handling, buffer pool lookups, page locking) constitute the majority of query execution time in a DBMS, often far exceeding the cost of distance computations.
• The optimal FVS algorithm choice is highly workload-dependent, influenced by filter selectivity, correlation, vector dimensionality, and result size, challenging library-centric assumptions that prioritize distance calculations.
• Clustering-based methods like ScaNN generally leverage sequential access patterns, SIMD vectorization, and improved cache locality to often outperform graph-based methods, particularly for lower-dimensional data.

Summary by alphaXiv.