Design and Implementation of a Scalable Scenario Query System for Meituan

Meituan, the largest lifestyle service platform, needs to support granular scene queries and intelligent suggestions for billions of users across millions of POIs. The system addresses three key problems: locating real‑time and historical traffic, understanding past activities in a target area, and recommending suitable investment locations.

The main challenges are massive data volume causing unacceptable real‑time response times, numerous RPC services whose latency accumulates, and the absence of a unified data model leading to high code coupling.

The solution adopts a layered architecture: bottom RPC services, a pre‑computation layer, a middle‑platform service layer, and business‑logic on top. The backend service layer provides Thrift RPC interfaces delivering raw feedback data.

Data assembly combines results from multiple services, filtering and intersecting list‑type responses while handling single‑value data, as illustrated in the original Fig 3.

To overcome assembly bottlenecks, the system pre‑computes and stores results. The pre‑computation layer builds a generic two‑level tree data model (aggregate root and geographic leaf nodes) and performs high‑density parallel computation using Apache Storm, enabling real‑time updates.

The tree model is persisted in a NoSQL KV store: Squirrel (Redis‑cluster) for fast writes and Tair for backup, keyed by ID+level to retrieve any node efficiently.

Two computation modes are compared: a serial layer‑by‑layer approach with high time and space complexity, and a parallel Storm topology that reduces both by processing each layer concurrently and streaming results downstream.

The middle‑platform service layer exposes unified RPC APIs for pagination, compression, filtering, batch extraction, and atomic retrieval, while integrating Rhino for circuit‑breaker and fallback mechanisms.

After deployment, response times dropped from several seconds to under one second, coupling between data and code was significantly reduced, and development effort for future feature changes was minimized.