How Meituan Food Service SaaS Built a Data Middle Platform on StarRocks
This article describes how Meituan Food Service SaaS built a high‑quality, large‑scale data middle platform using StarRocks, covering business overview, technical selection, multi‑layer architecture, virtual views, intelligent tiered querying, multi‑active hot standby, and the performance gains achieved.
The article introduces Meituan Food Service SaaS and its data products, outlining the business scenarios such as operation analysis, intelligent decision‑making, business alerts, and financial reconciliation, which demand high data quality, large volume, fast iteration, and good query experience.
It then details the system architecture, which consists of five layers: data synchronization, data production (offline and real‑time warehouses), data storage (MySQL, later extended with TiDB), data services, and data applications. Additional monitoring and stability systems ensure data quality.
Business pain points include massive data volume (single tables up to 10 TB), high query complexity, and the need for independent data‑value extraction by large merchants, requiring heterogeneous data source integration and low‑cost deployment.
For technology selection, the team evaluated AP engines (Druid, ClickHouse, Kylin, StarRocks) and chose StarRocks based on criteria such as join capability, standard SQL support, low operational cost, and ability to support independent deployment.
The new data middle platform built on StarRocks retains the layered model but introduces virtual views, eliminating the need for separate offline/real‑time compute clusters and enabling low‑cost, high‑efficiency deployment. Virtual views push SQL down through layers until reaching a materialized layer for execution.
Intelligent tiered querying routes small‑scale queries to OLTP databases (MySQL/TiDB) and large‑scale queries to StarRocks, reducing OLAP concurrency pressure and improving query ROI. An SDK automatically selects the appropriate data source based on static rules and dynamic workload analysis.
Multi‑active hot‑standby adds resilience by switching between StarRocks primary/replica clusters and between OLAP and OLTP clusters based on data‑quality monitoring, with automatic downgrade and recovery mechanisms.
Performance results show a 28× query speed improvement, 0.16 qps per restaurant in the SaaS scenario, and significant latency reductions (tp90 +30%, tp99 +500%). The platform is now in pilot deployment, with plans for full production rollout.
DataFunTalk
Dedicated to sharing and discussing big data and AI technology applications, aiming to empower a million data scientists. Regularly hosts live tech talks and curates articles on big data, recommendation/search algorithms, advertising algorithms, NLP, intelligent risk control, autonomous driving, and machine learning/deep learning.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.