Understanding the Essence of Architecture: A Deep Dive into Weibo’s Large‑Scale System Design

Before discussing the essence of architecture, the author reflects on the sheer scale of a ten‑million‑user website, noting that strategic attention is required for such magnitude while tactical decisions may downplay it.

Using Uber as an example, the article shows that a daily order volume of one million translates to an average QPS of about 30, while a single backend server can handle 800‑1000 QPS, highlighting the need to consider data storage and traffic amplification.

Architecture is described as an abstract "rack" that holds business logic and algorithms, emphasizing abstraction, repeatable patterns, and forward‑looking design based on experience and industry foresight.

The author lists three core abilities for architects: abstraction (deduplication and reusability), classification (decoupling services and defining interfaces), and algorithmic performance optimization (CPU, memory, I/O, network).

Examples of architectural techniques include MySQL sharding, CDN acceleration, service‑oriented design, and message queues for asynchronous decoupling.

Weibo’s overall architecture is presented as a three‑tier system: client layer (Web, Android, iOS), an interface layer with three main roles—security isolation, traffic control, and platform differentiation—and a backend layer consisting of platform services, search, and big‑data processing.

Key design principles are outlined: RPC services, messaging middleware for async processing and traffic shaping, and configuration management for gray releases and system degradation.

The importance of statelessness at the interface layer is discussed, explaining that state is shifted to caching or storage layers to maintain scalability.

Data‑layer design is emphasized as more critical than service‑layer design; poor schema decisions can lead to costly migrations and performance degradation.

Physical‑to‑logical mapping is illustrated, showing how teams are organized into vertical business groups and a core infrastructure group to improve communication.

The article details the request flow from DNS resolution to load balancer (VIP) before reaching the interface layer, underscoring hidden latency sources.

Performance bottlenecks in distributed systems ultimately reside in CPU, memory, storage, or network, and diagnosing them requires thorough monitoring.

Weibo’s multi‑level cache architecture (L1 and L2) is explained: L1 boosts QPS and bandwidth for hot content, while L2 handles capacity for massive data sets, reducing database penetration.

CDN and hybrid local‑plus‑distributed caching strategies are presented as common multi‑level cache solutions.

Feed storage uses MySQL with primary and secondary indexes, sharding by user ID and time to separate hot and cold data, preventing table explosion.

Distributed tracing is introduced to correlate logs across services using a unique request ID, enabling end‑to‑end visibility with minimal intrusion via AOP or similar techniques.

Capacity planning involves defining SLA metrics (CPU, memory, QPS) and monitoring real‑time traffic to trigger alerts.

Operational practices for handling traffic spikes include graceful degradation, full‑stack load testing (amplifying traffic 5‑10×), and shared Docker clusters for resource efficiency.

The article concludes with a personal learning roadmap: mastering Java, JVM, operating systems, design patterns, TCP/IP, distributed systems, data structures, and algorithms.