What Really Drives Scalable Backend Architecture? Lessons from Weibo’s Massive System

Architecture and My Understanding of Its Essence

Before discussing the essence of architecture, consider the magnitude of a ten‑million‑user website: Uber processes about a million orders daily with an average QPS of 30, while a single backend server can handle 800‑1000 QPS. The data storage, order routing, and downstream services quickly scale to billions of requests.

Architecture, as I once read on Zhihu, is essentially a framework that abstracts repetitive business logic and anticipates future expansion, much like a clothes rack that holds various items.

Key Skills for an Architect

Abstraction: eliminate duplication across functions, classes, services, and templates to improve reusability.

Classification: decouple objects, define service boundaries, and create modular interfaces.

Algorithm & Performance: optimize CPU, memory, I/O, and network.

Typical Architectural Patterns

Examples include MySQL sharding (using templates for routing), CDN for network acceleration, service‑oriented design, and message queues for asynchronous decoupling.

Weibo’s Overall Architecture

Weibo follows a three‑tier model: client (Web, Android, iOS), an interface layer, and backend services. The interface layer provides security isolation, traffic control, and separates PC and mobile needs. Backend services are divided into platform services, search, and big‑data processing.

As traffic grows from millions to billions of users, the architecture evolves through three generations, each requiring more sophisticated service decomposition and stability.

Design Principles

RPC services for remote calls.

Message middleware for asynchronous communication and traffic smoothing.

Configuration management for gray releases and graceful degradation.

Stateless interfaces shift state to the data layer, while careful data‑layer design prevents costly schema migrations.

Multi‑Level Caching

Weibo uses a dual‑level cache (L1 and L2). L1 handles extreme QPS spikes with low‑cost scaling, while L2 provides larger capacity to reduce database penetration. Similar patterns appear in CDN and hybrid local‑plus‑distributed caches.

Distributed Tracing and Monitoring

When services reach tens of millions of requests, full‑link tracing with unique request IDs becomes essential for pinpointing failures across RPC calls, especially in heterogeneous language environments.

Operational Practices

During peak events (e.g., Chinese New Year), Weibo employs degradation plans, full‑stack load testing, and a shared Docker cluster to ensure resilience.

Continuous Learning Path

Master Java, then the JVM, operating systems, design patterns, TCP/IP, distributed systems, data structures, and algorithms to build a solid architectural foundation.

Conclusion

Architecture is an evolving discipline; by constantly learning, practicing, and summarizing, you can develop your own principles for designing robust, scalable backend systems.