TiDB, One Layer at a Time: How Distributed SQL Became an Agentic AI Backbone

Every distributed database is a stack of hard problems solved in order, and the order tells a story. This talk traces how TiDB grew, layer by layer, from a single idea into a system now serving agentic AI platforms. It began over a decade ago with one job: speak SQL, but scale beyond a single machine. That demanded a distributed storage layer that could grow writes and capacity horizontally, and then a brain to keep that storage balanced, available, and consistent as machines come and go. Three layers, three clean responsibilities: a MySQL-compatible SQL engine, a Raft-replicated key-value store, and a placement driver. That separation is why the system kept absorbing new use cases, real-time analytics, change streaming, Vector and full text search, each as a new layer rather than a rewrite. Then the twist: none of this was designed for AI. Yet the very properties that make a good distributed database, elastic scale, strong consistency, fresh data, a clean interface, turn out to be exactly what AI agents need from their data backbone.