Tencent News Recommendation Architecture Upgrade: From Legacy Systems to a Scalable AI-Driven Platform

Programmers find great satisfaction when their code runs on millions of devices, but maintaining legacy code for such a massive user base is a major challenge. Tencent News, a ten‑year‑old product with hundreds of repositories and millions of lines of code, faced issues of poor availability, low scalability, and complex operations as it moved from a portal model to a recommendation‑driven model.

The article first reviews the business history of Tencent News, highlighting its rapid growth, peak DAU of 250 million in 2014, and the subsequent need to upgrade its recommendation architecture. It outlines three development stages: an early category‑based content showcase, a personalized waterfall stage driven by recommendation algorithms, and the current phase requiring a unified, high‑performance system.

Key problems of the old architecture included sub‑99% availability, long development cycles, high operational costs, and difficulty tracing issues across more than 200 repositories and diverse protocols.

The upgrade goal is to retain business logic independence while improving robustness, scalability, and operability. The solution adopts a domain‑driven design, consolidates data assets (users, content, features, strategies), and builds a unified data platform to support both personalized recommendation and push scenarios.

For the indexing layer, the new design replaces batch processing with streaming updates, introduces sharding and a multi‑chain inverted index, and selects a doc‑hash partitioning strategy to achieve high performance and low latency. Performance gains include up to 62% CPU reduction and 30% memory savings.

The feature platform is re‑engineered to provide lifecycle management, unified online services, and a clear separation between data collection and extraction. Centralized feature extraction, operator‑based pipelines, and consistent metadata ensure feature consistency and dramatically reduce extraction time.

A dedicated debugging platform is introduced to collect fine‑grained data, provide end‑to‑end traceability, and enable rapid issue diagnosis across the entire recommendation pipeline. This platform improves data freshness, supports minute‑level metric updates, and offers a unified view for developers, product, and operations.

Stability measures include request anomaly detection, graceful degradation, rapid hotspot scaling, comprehensive monitoring, and strict release processes (code coverage, CI/CD, canary releases). These practices raise system availability from below 99% to 99.99% and cut overall cost by more than 60%.

In conclusion, the continuous architectural evolution of Tencent News demonstrates how systematic redesign, data unification, and robust operational practices can sustain massive, fast‑changing services while delivering better performance, lower cost, and higher reliability.