How Huawei’s Multi‑Master Cloud Database Beats Aurora and CockroachDB

01 Introduction

Traditional high‑performance databases use a single‑writer (master) and many readers (replicas), which creates a single point of failure and limits write scalability. Multi‑master databases address these issues, and they can be built with either Shared‑nothing or Shared‑storage architectures. The paper discusses the challenges of Shared‑nothing under unbalanced workloads and the unsuitability of Shared‑storage for cloud‑native environments.

02 Huawei Cloud Database Multi‑Master Innovations

Huawei’s solution, presented in the VLDB 2023 paper “Taurus MM: bringing multi‑master to the cloud”, reduces network traffic in three ways:

Log‑as‑Database: writes are logged and replayed on storage nodes, avoiding full page writes over the network.

Vector‑Scalar (VS) Clock: combines scalar timestamps for frequent messages with vector timestamps for rare global snapshots, cutting message size.

Hybrid Row‑Page Locking with a Global Lock Manager (GLM): row‑lock information is piggy‑backed on page locks, and locks are managed locally when possible, dramatically lowering lock‑related traffic.

These techniques are illustrated in Figures 1 and 2.

03 Performance Evaluation

The authors benchmarked the system on a cluster with up to eight master nodes and four storage nodes (25 Gbps network, 128 GB buffer pool). Workloads included Sysbench (write‑only, 80 % read/20 % write) and TPC‑C with varying data‑sharing percentages.

Results show near‑linear scalability for partitioned workloads: with eight masters, 10 % shared Sysbench write achieved 3.5× speed‑up, read‑write 4.5×, and TPC‑C 5× compared to a single master. Scalability degrades as sharing increases (e.g., 30 % shared write <2×).

Compared with Amazon Aurora multi‑master, Huawei’s system consistently outperformed Aurora in both throughput and latency, especially under low‑share write workloads where Aurora suffered many aborted transactions.

Against CockroachDB (shared‑nothing), Huawei’s multi‑master delivered 60‑320 % higher throughput and significantly lower latency across 6‑node and 12‑node configurations.

04 Conclusion

Huawei Cloud Database Multi‑Master is a cloud‑native OLTP system that separates compute and storage, uses a global lock manager, and introduces VS clocks and hybrid row‑page locks to cut network overhead. Experiments confirm high performance and scalability, achieving up to 84 % efficiency on four masters and 62 % on eight masters in TPC‑C, and outperforming both Aurora and CockroachDB under typical workloads.