Design and Implementation of a Billion‑Scale Generalized Recommendation System at Tencent Cloud

Recommendation systems have become integral to daily digital experiences, influencing music, shopping, video, and news consumption. They improve user experience, boost product sales, uncover long‑tail value, and streamline mobile interactions.

Use Case Example : In the App Store, different users receive personalized app suggestions based on their preferences, reducing search time and increasing conversion rates.

To support over 200 recommendation scenarios across 12 Tencent services handling hundreds of billions of daily requests, two core challenges must be addressed: supporting numerous business scenarios and handling massive online traffic.

1. Generalized Recommendation Algorithm Library

A reusable algorithm library enables the same algorithms to serve multiple scenarios, avoiding duplicated development. The system consists of four components: a sample library, a feature library, algorithms, and models. Samples store user behavior logs, features capture item and user attributes, algorithms train models, and models are the trained outputs. An algorithm‑configuration table decouples algorithms from specific samples and features, allowing reuse.

2. Real‑Time Computing Platform for Massive Online Services (R2)

R2 is a custom stream‑processing framework inspired by Apache Storm, designed for high availability, plug‑in compute resources, dynamic scheduling, and integrated monitoring. It processes billions of recommendation requests daily with an average latency of 18 ms and 99.99 % stability.

Typical request flow: retrieve user features by ID, apply a decision‑tree model to estimate app preference probabilities, and re‑rank results using parallel subtree computation.

The R2 architecture comprises three layers: business (topology and processing units), communication (named messaging, stream handling, PU coordination), and global configuration (topology mapping and name service via Zookeeper).

3. Tencent Cloud Recommendation Engine (CRE)

Built on the above foundations, CRE offers a one‑stop, cloud‑based recommendation solution for small‑ and medium‑size internet businesses, providing secure, convenient, precise, and reliable services that improve click‑through rates and user experience.

Key features include one‑click onboarding, template‑based algorithms that reduce code by 99 %, rapid scaling, and stable operation that lowers operational costs.

Conclusion

To build a hundred‑billion‑level generalized recommendation system, one must design a reusable algorithm library, deploy a low‑latency, scalable, and reliable real‑time platform, and provide an easy‑to‑integrate cloud recommendation solution.