Agentic AI Architecture for Scale: A Unified Database on TiDB

Description

In this TiDB SCaiLE 2025 opening keynote, PingCAP Co-Founder and CEO Max Liu explores how agentic AI changes both the exploration and scale-up phases of successful product development, and what infrastructure you need when agentic AI meets real-world growth.

You’ll discover the two-stage journey every successful product faces: Stage 1 (explore rapidly, fail fast) and Stage 2 (mass adoption with real users, real data, and real growth pains). Max argues that agentic AI shines when you move beyond sequential prompts into parallel exploration. Think “Git for innovation,” where you branch ideas, test many paths at once, and dramatically increase the odds of finding what works.

Max shows how TiDB removes bottlenecks when that exploration hits traction: petabyte-scale storage, millions of QPS/TPS, millions of tables/schemas per cluster, and millions of connections. TiDB also enables large-scale batch ingest, online DDL concurrency, backup/restore, vector search for retrieval, resource groups, and query plan binding. Recent gains include 3M tables (3× year-over-year), 50× faster table creation, indexing at ~1M rows/sec, and 10× throughput for cross-region replication. This makes TiDB a strong fit for multi-tenancy deployments that agentic AI platforms often require.

A real-world example: Dify, an open-source agentic AI platform, consolidated structured, vector, and document data into a single TiDB offering. The company shipped 4.6× faster and cut database costs 80% versus managing per-tenant instances. This shows how agentic AI platforms benefit from a unified data plane as they scale from prototype to production.

You’ll learn:

How agentic AI accelerates Stage-1 exploration with parallelism
The database features that unlock Stage-2 scale without rewrites
Why multi-tenant design matters for agentic AI platforms- Practical metrics and improvements to watch (DDL throughput, replication, and indexing)