Evolution of Meituan Delivery System Architecture and Practices

Since its inception, Meituan Delivery has undergone rapid growth, demanding continuous upgrades to its system architecture and infrastructure. This article shares the team’s reflections and practices for supporting high availability, scalability, and development efficiency in a large‑scale, real‑time delivery platform.

Business context : The rise of e‑commerce and O2O models (food, groceries) has driven the emergence of same‑city instant delivery, characterized by fast timeliness, short distances, and high randomness in pickup and drop‑off locations.

Technical challenges : The platform must satisfy stringent online SLA requirements while handling complex offline operations across multiple business lines.

Architecture evolution is divided into three stages:

MVP stage : Rapid experimentation with coarse‑grained service decomposition based on three domains – people, finance, and assets.

Scaling stage : Focus on overall architecture evolution, fulfillment system availability, and operational efficiency, adopting a “simplify‑then‑divide‑then‑evolve” strategy.

Fine‑grained stage : Mature business models and refined operations, emphasizing AI‑driven optimization, simulation platforms, and feature pipelines.

The system is split into three core subsystems:

Fulfillment system : Handles order creation, payment, and dispatch, separated for user‑side and rider‑side concerns.

Operation system : Manages planning, business management, rider operations, and settlement, providing unified workflows and approvals.

Master Data Platform : Centralizes core entities (organizations, personnel, merchants, cities) and supports CQRS/MDM patterns across production, core model, capacity, and management services.

Reliability engineering follows a pre‑/mid‑/post‑incident framework, building preventive processes, diagnostic tools, and automated recovery mechanisms. Disaster‑recovery progresses from dependency degradation to end‑to‑end fault isolation, with real‑time monitoring of key metrics (order volume, online riders) and trace‑based root‑cause analysis.

Capacity planning combines static assessments, trace‑based monitoring, and large‑scale load testing using traffic tagging, shadow tables, traffic shifting, and replay to evaluate and expand system capacity through redundancy, vertical/horizontal sharding, and automated archiving.

Operational efficiency is improved via dynamic forms, workflow platforms, and standardized components, reducing manual effort and accelerating feature delivery.

AI and simulation : A sandbox environment replays order flows, models user/merchant/rider entities, and simulates rider behavior to evaluate algorithmic KPI offline. An algorithm data platform provides end‑to‑end pipelines for data cleaning, feature extraction, model training, and online inference.

LBS platform : Supplies location‑based services (mapping, routing, heatmaps) that underpin dispatch, ETA prediction, and pricing.

In conclusion, the architecture continuously balances ROI, technical innovation, and business growth, turning complex problems into manageable components and laying a foundation for future scalability and AI‑driven enhancements.