Chapter 12: Design a Chat System

A crucial decision in this design is selecting the proper network protocol for server-initiated updates, which is essential for low delivery latency; while traditional HTTP is adequate for the sender, techniques like polling and long polling are inefficient for the receiver side due to their periodic nature and potential server statelessness. The industry standard solution adopted is WebSocket, which provides a persistent, bi-directional communication connection that begins as an HTTP request and is subsequently "upgraded," simplifying the architecture for both sending and receiving clients. The system is architecturally segmented into three major categories: stateless services (handling features like login and user profiles), stateful services (the Chat Service maintaining persistent client connections), and third-party integration (primarily for push notifications). The data layer features a hybrid storage model, using relational databases for generic data (like user profiles) and highly scalable, low-latency Key-Value (KV) stores (such as HBase or Cassandra) for the enormous volume of chat history, as relational databases struggle with the "long tail" of older, less-frequently accessed data. To guarantee message sequence, a unique, time-sortable message ID must be generated, often accomplished using a local sequence number generator. During deep dives into system components, the role of Service Discovery (e.g., Zookeeper) is highlighted as coordinating server capacity and location to recommend the best chat server to a client upon login. The messaging flow ensures that every message is assigned an ID, stored in the KV store, and routed via a Message Synchronization Queue to the recipient, triggering a push notification if the recipient is offline. Furthermore, maintaining Online Presence relies on dedicated Presence servers that utilize a heartbeat mechanism—where clients periodically ping the server—to accurately determine if a user is online, thereby preventing poor user experience from frequent status changes caused by minor network interruptions, and status updates are efficiently distributed to friends using a publish-subscribe model.