X‑Stor: A Cloud‑Native Multi‑Model NoSQL Database Service – Design, Architecture, and Evaluation

Paper Background – VLDB 2024 selected the paper "X‑Stor: A Cloud‑native NoSQL Database Service with Multi‑model Support" from Tencent X‑Stor team and Huazhong University of Science and Technology. The work addresses the challenges of managing many heterogeneous NoSQL systems in large‑scale cloud environments.

Motivation – Using separate NoSQL engines for different workloads leads to duplicated development, operational complexity, and resource waste. X‑Stor is built to provide a unified, cloud‑native platform that supports KV, time‑series, feature storage and other models while offering strong, eventual, and bounded consistency.

System Overview – X‑Stor runs on Tencent Kubernetes Engine (TKE) with full containerisation, offering elastic scaling, high availability, and multi‑AZ deployment. The architecture separates a control plane (resource allocation, scheduling, scaling, health‑checking) from multiple data‑plane instances, each dedicated to a specific data model.

Multi‑Model Design – X‑Stor adopts a plugin‑based storage engine layer: each model (KV, time‑series, feature) is served by a lightweight engine exposing only basic data‑access interfaces, while auxiliary components (WAL, replicator, etc.) are deployed separately. Adding a new model only requires implementing a new plugin, reducing development cost and allowing independent optimisation.

Process‑Level Multi‑Tenant Support – Tenants share a single DBMS process, reducing resource fragmentation. X‑Stor introduces a unified metric called Request Unit (RU) that normalises CPU, memory, IOPS and network consumption of a request across different models, enabling fair isolation, serverless‑style billing, and hardware‑independent accounting.

Experimental Evaluation

Performance tests compare X‑Stor’s three plugin engines (LSM‑tree, in‑memory, time‑series) against LevelDB, Redis and InfluxDB using db_bench, YCSB and TSBS. Results show X‑Stor consistently outperforms the single‑model counterparts.

Tenant isolation experiments using the RU metric demonstrate that, under increasing resource contention, RU‑based isolation keeps latency lower than a naïve setup, proving its effectiveness for heterogeneous workloads.

Load‑balancing tests on a production cluster show that RU‑driven pod migration eliminates overloaded pods, reduces standard deviation of load across pods, and slightly improves overall request latency.

Overall, X‑Stor demonstrates that a cloud‑native, plugin‑based, multi‑model NoSQL service can achieve high performance, fine‑grained resource isolation, and efficient scaling, making it suitable for diverse internal and external workloads.