Alkin Tezuysal

Vector Search Wars: Competing Architectures in Open Source Databases

AI has sparked the latest evolution in database technology. In just a few years, vector search has gone from a specialized capability to a core feature across the open source database ecosystem. Today, MySQL, MariaDB, PostgreSQL, ClickHouse, and TiDB all support vector search, but each has taken a fundamentally different architectural approach. This session explores the engineering decisions behind these implementations rather than comparing feature lists or benchmark numbers. We'll examine how each database integrates vector data into its storage engine, the trade-offs between native implementations and extensions, the use of approximate nearest neighbor (ANN) indexes such as HNSW and IVF, and the impact of filtering, hybrid search, and distributed execution on real-world workloads. We'll also discuss how emerging standards such as the Model Context Protocol (MCP) are enabling AI applications and agents to interact with databases, and how these interfaces complement existing vector search architectures without replacing them. Rather than declaring winners and losers, this talk explains why these databases made different design choices, which workloads those choices favor, and where the open source ecosystem is converging. Whether you're building semantic search, Retrieval-Augmented Generation (RAG) applications, AI assistants, or recommendation systems, you'll leave with a practical framework for evaluating vector search capabilities based on architecture, scalability, operational complexity, and long-term maintainability.