Design and Evolution of Ctrip's Hermes Message Queue System

Gu Qing introduces Hermes, the new message system being deployed at Ctrip, focusing on its architecture and implementation details.

The advantages of message queues are highlighted, including decoupling of services, asynchronous processing, handling traffic spikes, and supporting fan‑out scenarios.

The basic MQ model is explained with queue and topic patterns, consumer groups, and message delivery semantics.

Version 1.0 of Ctrip's MQ stored messages directly in MongoDB without a broker, leading to high client upgrade costs, heavy DB coordination, limited features, and poor elasticity.

Version 2.0 added a master‑slave broker layer that coordinates via MongoDB heartbeats, allowing clients to communicate only with brokers, simplifying upgrades and reducing client complexity.

The current architecture (Fig 4) consists of producers, brokers, and storage back‑ends (MySQL for critical data and Kafka for high‑throughput logs), with a meta‑server handling cluster coordination and lease management.

Two message storage types are used: Kafka provides high throughput but lacks features like replay and priority, while MySQL offers richer queue capabilities for important business data.

Efficient MQ construction starts with single‑machine optimizations: fast insert‑only tables, batch writes, minimal indexing, and low‑latency delivery via push or long‑polling pull mechanisms.

Scaling to a cluster involves adding brokers, partitioning topics, and ensuring ordering within partitions; load balancing is achieved by assigning partitions to specific brokers.

Cluster management relies on a lease‑based approach where the meta‑server grants time‑limited leases to brokers and consumers, using ZooKeeper for broker coordination and HTTP/TCP protocols for meta‑server communication, enabling simple HA and dynamic rebalancing.

The summary emphasizes the need for fast writes, large throughput channels, low delivery latency, partition stickiness, long‑pulling techniques, and lease‑based coordination to build a robust, scalable message queue system.