Meituan Instant Logistics: Distributed System Architecture and Practices

Meituan's instant logistics has grown over five years, accumulating experience in building distributed high‑concurrency systems that must tolerate near‑zero failures and low latency while handling massive order and rider scales.

The platform focuses on three core goals: providing SLA guarantees such as ETA and pricing, matching riders under cost‑efficiency‑experience trade‑offs, and offering rider‑side decision support like voice interaction and route recommendation.

To support these goals, Meituan evolved from vertical services to layered services and finally to micro‑services, emphasizing gradual evolution rather than premature design.

Key components of the distributed architecture include HLB for front‑end load balancing, OCTO for service registration, discovery, load balancing and fault tolerance, message queues (Kafka, RabbitMQ), Zebra for distributed database access, CAT for monitoring, Squirrel+Cellar for caching, and Crane for task scheduling.

Scalability challenges such as stateful node expansion and hotspot mitigation were addressed by converting stateful nodes to stateless ones and leveraging parallel computation for rapid scaling.

Data consistency across databases and caches is ensured by Databus, a high‑availability, low‑latency pipeline that captures binlog changes and propagates them to downstream stores like Elasticsearch.

High availability is maintained through pre‑emptive full‑stack stress testing, periodic health checks, chaos engineering (service, machine, component failures), real‑time anomaly alerts (performance, business metrics, availability), rapid fault isolation (single‑machine, cluster, IDC, component, service), and post‑incident rollback, throttling, circuit‑breaking, and fallback mechanisms.

Deployment strategies include rapid single‑IDC failover with automatic traffic switching, pre‑synchronised data and services for quick scaling, and multi‑IDC virtual centers that treat groups of IDC as a single partition for elastic capacity.

Unit‑based deployment further refines disaster recovery and scaling by routing traffic based on regions or cities, handling cross‑region data latency, and enabling swift SET failover.

AI capabilities are consolidated in a machine‑learning platform for model training and deployment, and the JARVIS AIOps platform improves incident handling by de‑duplicating alerts and automating fault analysis.

Future challenges include managing the complexity of expanded micro‑services, mitigating network amplification from latency, accelerating fault localisation in intricate topologies, and transitioning operational practices from cluster‑level to unit‑level management.