Ensuring Data Consistency in Meituan Hotel Direct Connection Platform

Meituan's hotel direct‑connection project started at the end of 2013 and has grown to integrate more than 200 suppliers, over 30,000 online hotels and 300,000 SPUs. By the end of 2015 it became the largest hotel‑direct platform in China, expanding from budget chains to high‑star channels, small groups and non‑standard accommodations.

With the rapid growth of Meituan‑Dianping, the platform faces increasing traffic and data volume, which puts pressure on its technical architecture and data‑application layers. The team must maintain high availability, fast response times, and data consistency while keeping operational risk low.

Problem Statement : How to keep product cache and supplier data consistent during the order‑placement process, especially when supplier interfaces are unstable?

The system must transform supplier hotel products (room types) into Meituan‑Dianping sellable items, handling price, inventory, and cancellation rules for a 60‑day booking window. This requires reliable synchronization of large‑scale product data.

Overall Idea : Treat the consistency challenge like MySQL master‑slave replication. MySQL uses three threads (Binlog Dump on the master, I/O thread on the slave, and SQL thread on the slave) to copy binary logs and apply changes. The same principle can be applied to push supplier changes to the direct‑connection platform.

Solution Evolution – Four Stages :

Stage 1 – From Nothing to Everything : Periodically pull the full product catalog from suppliers. Advantages: short development cycle, simple logic. Disadvantages: high bandwidth and storage demand, long recovery time, and data may be stale for up to 30 minutes.

Stage 2 – Divide and Conquer : Pull data in segments. Full 60‑day data is fetched at fixed times (e.g., 1 am, 7 am, 1 pm, 8 pm) while a 10‑day slice is refreshed every 15 minutes. This reduces traffic by ~5/6 and avoids duplicate task launches.

Stage 3 – Refine and Predict : Combine trigger‑based updates (validate product data just before payment) with inventory prediction. Hot products are pre‑updated based on user behavior, while less‑frequent items are pulled less often. Collaborative‑filtering or deep‑learning recommendation models can identify the top‑20 % of products that generate 80 % of revenue.

Stage 4 – Cooperative Push : Encourage suppliers to actively push data changes via standardized APIs. This reduces pull traffic, improves success rates, and lowers operational costs. Meituan provides a portal with documentation, testing tools, and analytics to facilitate integration.

Summary : By iteratively improving data acquisition—from full batch pulls to selective segment pulls, trigger‑based validation, predictive caching, and finally supplier‑initiated pushes—the platform achieved reliable, low‑latency product data for users. The final architecture combines passive pulling, active pushing, and real‑time verification to ensure consistency.

References include CSDN articles on MySQL replication and item‑based collaborative filtering, as well as several WeChat posts on recommendation systems and large‑scale data processing.