Skip to main content
← User Stories
JD · China

Connecting online and offline samples for model training

JD combines Fluss and Apache Hudi to produce recommendation-ranking samples for online serving and near-real-time and offline model training.

About JD

JD is a supply-chain technology and services company with businesses in retail, logistics, and related services.

A shared foundation for training samples

JD builds its recommendation-ranking sample pipeline on Apache Fluss and Apache Hudi. The architecture connects streaming and batch sample generation with model training, supporting online sample serving alongside near-real-time and offline training data.

Tiering samples into the lake

The pipeline uses the tiering service to synchronize data to Hudi lake storage in near real time. JD reports contributing the Fluss–Hudi integration back to the open-source community.

Reading only what a training job needs

During online training, jobs use Fluss column pruning to read only the fields needed by the model. This improves reading efficiency and reduces bandwidth consumption. JD also collaborates with the community on message-queue capabilities such as offset management.

Explore more user stories →