How Meituan Scaled Instant Delivery with Distributed Architecture and AI

Background

Meituan's food‑delivery service has operated for five years, and its instant logistics capability has been explored for over three years. The team accumulated experience in building high‑concurrency distributed systems. Two main takeaways are the extremely low tolerance for failures and latency, and the need to combine AI with cost, efficiency, and user‑experience optimization.

Massive order‑rider scale creates ultra‑large matching computation problems.

Holiday or severe‑weather spikes cause traffic to surge to many times the normal level.

Logistics fulfillment links online and offline, requiring near‑zero downtime and ultra‑high availability.

Real‑time data must be accurate and highly sensitive to delay or anomalies.

Meituan Instant Logistics Architecture

The platform focuses on three core functions: (1) providing SLA guarantees such as ETA calculation and delivery‑fee pricing; (2) matching the most suitable rider under multi‑objective (cost, efficiency, experience) optimization; (3) offering rider assistance during fulfillment, including intelligent voice, route recommendation, and store‑arrival reminders.

Behind these services lies a distributed architecture built on Meituan's common components. Front‑end traffic is balanced by HLB. Within each partition, services communicate via OCTO, which supplies service registration, discovery, load‑balancing, fault tolerance, and gray‑release capabilities. Messaging can also use Kafka or RabbitMQ. Storage accesses a distributed database through Zebra. System and business logs are collected, reported, and monitored by CAT, Meituan's open‑source distributed monitoring system. Distributed caching uses a Squirrel + Cellar combo, and task scheduling is handled by Crane.

Distributed System Practice

Typical challenges include cluster scalability, especially for stateful clusters, and hotspot resources such as uneven CPU usage.

Stateless‑ification: Convert stateful nodes to stateless, leverage parallel computation, and let small business nodes share load for rapid scaling.

Consistency: Use Databus, a high‑availability, low‑latency, high‑concurrency change‑data‑capture system, to propagate binlog changes from the primary DB to Elasticsearch, other DBs, or KV stores, ensuring eventual consistency.

High Availability: Conduct full‑link load testing to estimate peak capacity, perform periodic health checks, run random fault‑injection drills, set up performance and business‑metric alerts, enable fast fault localization (single‑machine, cluster, IDC, component, service), collect change logs before/after incidents, and apply rollback, rate‑limiting, circuit‑breaking, degradation, and fallback mechanisms.

Rapid Deployment & Disaster Recovery for a Single IDC

When a single IDC fails, entry services detect the fault and automatically switch traffic. Fast IDC expansion is achieved by pre‑synchronizing data, pre‑deploying services, and opening traffic only after the services are ready. All data‑sync and traffic‑distribution services must support automatic fault detection, removal, and IDC‑level scaling.

Multi‑Center Experiment

Because a single partition may hit resource limits, Meituan groups several IDC into a virtual center, treats the center as a partition unit, deploys services uniformly across the center, and adds new IDC to expand capacity when needed.

Unit‑Based Experiment

Compared with the multi‑center approach, unit‑based design offers a better solution for partition disaster recovery and scaling. Traffic routing is based on region or city characteristics. Cross‑region data synchronization may introduce latency. SET disaster recovery ensures that if a local or remote SET fails, traffic can be quickly switched to another SET.

Core Technical Capabilities and Platform Foundations of Smart Logistics

The machine‑learning platform provides an end‑to‑end pipeline for offline‑to‑online model training and algorithm deployment, solving the problems of repeated work and inconsistent data quality between offline and online sources.

JARVIS is an AIOps platform aimed at stabilizing operations. It aggregates noisy alerts, reduces duplicate alarms, and automates fault analysis to improve response speed and reliability.

Future Challenges

After review, several challenges remain: microservices are no longer "micro" as business complexity grows; mesh‑style service clusters amplify even slight latency; complex topologies make rapid fault localization difficult; and moving from cluster‑level to unit‑level operations will heavily test Meituan's deployment capabilities.