How Tencent’s X‑Stor Tackles Multi‑Model NoSQL at Cloud Scale

Read: Tencent uses “X‑Stor” to control database sprawl and save cloud resources.

Chinese internet giant Tencent unveiled a proprietary NoSQL database called X‑Stor, which it claims handles multiple data models more elegantly than existing solutions and improves resource utilization across its massive services.

The system was disclosed in a paper published in the VLDB Journal, a nonprofit venue for database research.

Paper link: https://www.vldb.org/pvldb/vol17/p4025-lei.pdf

The paper notes that NoSQL databases are typically built for specific data models. Tencent runs several databases to support its social network, video streaming, online gaming, and public cloud services, serving over a billion active users.

According to the paper “X‑Stor: A Cloud‑Native Multi‑Model NoSQL Database Service”, Tencent uses a graph database for social relationships, a wide‑column store for user profiles, a document store for its advertising platform, and a time‑series database for user behavior data.

However, this fragmented approach proved suboptimal; adding new data models required building new NoSQL systems from scratch, leading to duplicated effort.

Large‑scale deployments of heterogeneous databases cause resource isolation, complicating maintenance and hindering effective resource sharing.

X‑Stor addresses this by extending its storage engines and data access interfaces to support various models, allowing each engine to perform comparably to single‑model engines.

The authors claim X‑Stor is more elegant than competing NoSQL solutions such as MongoDB, Redis, and ArangoDB.

X‑Stor runs as serverless micro‑services orchestrated by Tencent’s Kubernetes Engine.

Initially deployed on SSD‑equipped hosts to meet I/O‑intensive workloads, X‑Stor later optimizes node selection to match workloads with appropriate resources, maximizing utilization.

The paper includes detailed mathematical models describing workload competition and resource allocation.

X‑Stor can scale to store 12 PB of operational data, handle 700 billion requests per day, peak at 30 million requests per second, and manage over 100 k tables with diverse data models.

Currently the system is not open‑source, limiting external experimentation.

Chinese internet giants continue to innovate in large‑scale data infrastructure.