Evolution of the DaJia Recommendation System Architecture: From V1.0 to V3.0

Recommendation has become a core competitive factor for e‑commerce platforms, appearing in many pages such as home, detail, cart, order success, and error pages, each with different focus.

The recommendation system improves user experience and helps businesses address long‑tail and Matthew effects, increase user stickiness, and boost product value and profitability.

The "DaJia" recommendation system has undergone three architectural versions.

V1.0 used a simple strategy‑plus‑factory design, enabling rapid business iteration.

V2.0 vertically split the system by business scenario and horizontally divided the pipeline stages, improving isolation, scalability, and resource allocation.

V3.0 modularized each recommendation stage, introduced a configuration service for dynamic pipeline configuration, and separated compute‑intensive prediction and I/O‑intensive recall into independent services.

V1.0 faced problems such as poor fault isolation, resource contention, limited scalability, increasing complexity with business growth, Redis bottlenecks, and risk of a single Redis cluster.

V2.0 addressed these by vertically splitting services per business, sharding Redis clusters, and adopting a configurable pipeline that assembles handlers for stages like recall, filter, coarse ranking, merge, fine ranking, intervention, and dispersion.

V3.0 added a configuration server and client that dynamically manage pipeline details and experiments, enabling A/B testing at the handler level, reducing code coupling, and allowing real‑time adjustments.

The system also migrated the recall data store from Redis to Elasticsearch, introduced a prediction service supporting multiple models, and built an AB testing capability.

Future work aims to build an explanation platform for recommendations, improve real‑time feature handling, and achieve personalized, explainable recommendations for each user.