Design and Evolution of Meituan's Real-Time Logistics Distributed System

Background

Meituan's food delivery has grown for five years and its instant logistics for over three years, accumulating experience in building high‑concurrency distributed systems. The main takeaways are low latency tolerance and the need for scalable, fault‑tolerant architecture.

Meituan Instant Logistics Architecture

The platform focuses on three core tasks: providing SLA guarantees such as ETA calculation and pricing, matching riders under cost‑efficiency‑experience trade‑offs, and offering rider‑side decision support (voice interaction, route recommendation, store‑arrival reminders). It leverages Meituan's common components, HLB load balancer, OCTO service registry, Kafka/RabbitMQ messaging, Zebra distributed database, CAT monitoring, Squirrel+Cellar cache, and Crane scheduler.

Distributed System Practices

Key challenges include massive order‑rider scale, traffic spikes during holidays, ultra‑low fault tolerance, and strict real‑time data requirements. Solutions involve converting stateful nodes to stateless, using Databus for DB‑cache consistency, extensive fault‑injection testing, periodic health checks, rapid rollback, and mechanisms such as rate limiting, circuit breaking, and degradation.

High‑Availability Deployment & Disaster Recovery

Describes single‑IDC fault detection with automatic traffic switching, rapid IDC scaling with pre‑synchronised data and ready‑state services, multi‑IDC virtual centres for capacity expansion, and unit‑level isolation to improve scaling and fault isolation.

Intelligent Logistics Core Technologies

Introduces a machine‑learning platform for end‑to‑end model training and deployment, and the JARVIS AIOps platform for alarm de‑duplication, automated incident analysis, and improving system stability.

Future Challenges

Future issues include increasing service complexity that inflates micro‑service size, network amplification effects from minor latency, fast fault localisation in complex topologies, and the operational shift from cluster‑level to unit‑level maintenance.