Architecture Design Overview of Recommendation Systems

Recommendation systems are one of the most successful AI applications in the mobile Internet era.

From an architectural viewpoint, this article reviews key algorithm modules, focusing on the three layers—offline, near‑line, and online—through which these algorithms are deployed.

▊ Personalization vs. Complexity

Higher personalization drives increasing system complexity, requiring more algorithms, features, and data dimensions, which in turn raises the model’s VC dimension and improves user differentiation.

▊ Timeliness vs. Computation

Improving response speed and data freshness enhances user experience but incurs higher computational costs and system complexity.

▊ Time vs. Space

Balancing cache usage and real‑time computation helps decouple producers and consumers, improving performance while managing resource trade‑offs.

System boundary definition clarifies what the recommendation system is responsible for (inputs) and what it depends on externally (outputs), guiding functional and team scope.

Key external dependencies include data sources (user behavior, item data), platform tools (ML platforms, real‑time and batch processing), and services that provide inputs or consume recommendation results.

Three architectural layers are defined:

Offline layer: batch processing of large volumes of non‑real‑time data, producing models and features for downstream use.

Near‑line layer: processes real‑time data but does not serve requests directly; results are written to caches for the online layer.

Online layer: real‑time service handling user requests, performing candidate recall, ranking, business logic, and result assembly within milliseconds.

Each layer’s characteristics, typical tasks, and trade‑offs are detailed, illustrating how they complement each other to achieve both efficiency and effectiveness.

The article also presents architectural diagrams (offline logic, near‑line computation, online service flow) and a comparative table of the three layers.

In conclusion, engineers should consider architecture alongside algorithms and data, ensuring a holistic, system‑level design for enterprise‑grade recommendation systems.