Exploration and Practice of Tencent Cloud Native Database TDSQL-C Architecture

Databases are undergoing a transformation driven by changing user demands and the emergence of new storage technologies. Cloud‑native databases, such as Tencent Cloud Native Database, deliver better performance, availability, and reliability by offering database services in a cloud model. This article is derived from the talk “Tencent Cloud TDSQL‑C Architecture Exploration and Practice” delivered by Zhang Qinglin, Technical Director of Tencent Cloud Database, at the Tech TVP Developer Summit.

The presentation is divided into three parts: (1) the background, architecture, and current status of TDSQL‑C; (2) breakthrough innovations of TDSQL‑C; and (3) the future technical roadmap.

Traditional database deployments face several challenges: limited storage capacity on a single node, cumbersome scaling (adding and removing machines during traffic spikes), HA recovery latency, and reliability issues when local disks fail or when backup/restore operations interfere with active workloads.

To address these problems, TDSQL‑C adopts a compute‑storage separation architecture. Compute nodes handle transaction processing, locking, and DML, while the storage layer (HiStore) provides up to 1 PB of network‑attached storage. Replication uses Redo logs instead of traditional binary logs, allowing a primary node and up to 15 read replicas. The storage layer persists data based on Base pages combined with Redo logs, eliminating WAL‑induced performance jitter.

Key metric breakthroughs include massive storage capacity, intelligent auto‑scaling, million‑level QPS, compatibility with MySQL and PostgreSQL, and the ability to scale within seconds via snapshot + Redo‑log mechanisms. Serverless capabilities enable per‑use billing (charged every 5 seconds), rapid instance start/stop, and page‑level billing that charges only for the storage actually written.

The second breakthrough focuses on a two‑level cache: a local SSD/AEP cache that stores frequently accessed data, dramatically reducing network I/O. In IO‑bound scenarios, even a sub‑50 % cache hit rate yields exponential performance gains.

The third breakthrough provides invisible backup and ultra‑fast restore. Backups are snapshots of HiStore’s distributed file system, performed in parallel across storage cells and uploaded to COS. Restores replay Redo logs directly at the physical level, achieving GB‑scale recovery in seconds.

Future directions for TDSQL‑C include ultra‑minimal operations (intelligent optimizer adjustments, instant DDL, parallel index creation), low‑database development for petabyte‑scale analytics, business‑specific features such as field‑level encryption and multi‑region disaster recovery, and the integration of a column‑store engine to accelerate analytical workloads.

The speaker, Zhang Qinglin, is the Technical Director of Tencent Cloud Database, a MySQL kernel architect, and a key contributor to the MariaDB community, leading the development of the cloud‑native database kernel.