Evolution of Weibo Recommendation System Architecture: From Independent 1.0 to Platform 3.0

Weibo’s recommendation platform, a broadcast‑style social network, relies on a follow‑based mechanism to deliver short, real‑time content, making recommendation a core component for both user relationship building and content propagation.

The article first outlines the generic recommendation pipeline (candidate generation, ranking, strategy, display, feedback, and evaluation) and then describes three major architectural phases.

1. Independent 1.0 (2011‑2013) – Small, isolated projects built with Apache + mod_python, Redis, and custom C/C++ services (woo). The focus was rapid business delivery, resulting in simple, easy‑to‑implement systems but lacking feedback loops, unified data handling, and robust operations.

2. Layered 2.0 (2013‑2014) – Introduced a multi‑layered design comprising an application layer (common_recom_frame), a computation layer (WOO framework), a data layer (static via Lushan, dynamic via Redis, with RIN/R9 interfaces), and foundational monitoring/evaluation services. This stage achieved a complete recommendation flow, data‑first principles, and better algorithm support, though it still fell short of a fully dedicated recommendation stack.

3. Platform 3.0 (2014‑present) – Builds on 2.0, adding standardized interfaces, candidate generation modules (Artemis, item‑cands), and the EROS strategy platform for model training, feature selection, and online testing. The architecture now emphasizes algorithm‑centric design, abstracted pipelines, and tighter integration of candidate, ranking, training, and feedback components.

The concluding section reflects on the symbiotic relationship between business needs and technical evolution, offering practical advice on incremental architecture selection, open‑source tool promotion, and focusing on what not to build.