Design and Implementation of Meituan's Traffic Compass Data Warehouse for Hotel‑Travel Business

Background : After entering the "second half" of the internet era, Meituan‑Dianping, the world’s largest life‑service platform, needed a powerful tool—named "Traffic Compass"—to analyze massive user traffic and drive growth for its hotel‑travel business.

Challenges : The system had to handle millions of daily UVs, exponential growth of dimensions, strict latency requirements (≤3 s for daily, ≤5 s for monthly queries), extensible dimension management, flexible entry‑point tracking, and correct aggregation logic for weekly/monthly metrics.

Solution Idea : Adopt Kylin as the core analytical engine, prune dimensions early, layer the data pipeline to reduce coupling, and standardize entry‑point tagging for unified metric calculation.

Architecture (see Fig. 2): A‑layer (ODS) collects raw app logs; B3‑layer extracts hotel‑travel specific logs; B2‑layer enriches with common dimensions; B1‑layer builds wide‑table topics; the App layer provides business‑specific models; a View layer buffers between the App layer and Kylin cubes; cubes serve aggregated data to backend services; a Permission layer controls access; and the Front‑end layer presents results to users.

Public Dimensions : Unified tagging rules create reusable dimensions such as page type, page detail, and traffic entry, ensuring consistency across layers (Fig. 3).

Topic Wide‑Table Layer : Provides rich, extensible dimensions and standardized business metrics while keeping query latency low; uses JSON strings for flexible, overlapping category dimensions.

Data Application Layer : Implements queries via Kylin cubes, offering simple backend logic, shielding business rules, and supporting easy business expansion.

Backend Service Layer : Handles query routing and configuration; entry‑point configurations (Fig. 7) are permission‑controlled and dictate which metrics and dimensions are available per business line and platform.

Evaluation Metrics : Business depth/breadth (model demand volume), response speed (request‑to‑delivery time), data stability, query efficiency, data quality (missing/incorrect/inconsistent data), and resource consumption (storage and compute).

Conclusion : Traffic Compass has been deployed for two quarters across Meituan‑Dianping’s hotel‑travel lines, receiving positive feedback; the system will continue to be refined and expanded.