Evolution of Meituan Instant Logistics Distributed System Architecture and Practices

Meituan's instant logistics platform has grown over five years, accumulating extensive experience in building distributed, high‑concurrency systems. Two major takeaways are the need for ultra‑low latency and fault tolerance, and the integration of AI across pricing, ETA, dispatch, capacity planning, and monitoring to improve scale, experience, and cost.

The platform focuses on three core tasks: providing SLA guarantees such as ETA and pricing, matching riders under multi‑objective optimization, and offering rider‑side decision support through voice, routing, and store‑arrival reminders.

Architecturally, the system evolved from vertical services to layered services and finally to microservices, emphasizing gradual evolution rather than premature microservice design. The distributed architecture relies on Meituan's common components, using HLB for load balancing, OCTO for service discovery and fault tolerance, and message queues like Kafka or RabbitMQ for communication. Storage accesses Zebra, monitoring uses CAT, caching combines Squirrel+Cellar, and task scheduling is handled by Crane.

Key operational challenges include massive order and rider volumes, traffic spikes during holidays or bad weather, stringent availability requirements, and real‑time data accuracy. Solutions involve converting stateful nodes to stateless ones for rapid scaling, ensuring data consistency with a high‑availability Databus that streams binlog changes to downstream stores, and implementing comprehensive fault‑tolerance practices such as full‑link stress testing, periodic health checks, chaos engineering, real‑time alerts, rapid root‑cause analysis, and automated rollback, rate limiting, circuit breaking, and degradation.

Deployment strategies cover single‑IDC rapid provisioning and disaster recovery, where entry services detect failures and switch traffic automatically, and multi‑IDC (or multi‑center) approaches that treat groups of IDC as virtual partitions to enable seamless capacity expansion.

Beyond infrastructure, Meituan built a machine‑learning platform to unify model training and deployment, addressing data quality and iteration efficiency. The JARvis AIOps platform was introduced to reduce alert noise, automate fault detection, and improve incident response speed.

Future challenges include managing the growing complexity of microservices, mitigating network amplification effects from latency, enhancing rapid fault localization in dense service meshes, and transitioning operational practices from cluster‑level to unit‑level management.