Chanmama

Solution

Chanmama redesigned its recommendation pipeline around EMR Serverless Spark and Vector Search Service Milvus Edition so product data and embeddings could be processed as a managed, elastic workflow instead of a self-operated cluster. Periodic offline jobs extract product records from StarRocks, including IDs, titles, and descriptions, then use a machine-learning model service to convert product titles into vectors. Serverless Spark combines those vectors with product metadata and batch-loads them into Milvus, which becomes the low-latency similarity-search layer for recommendation and creator product-selection workloads. Because the compute layer is serverless, the team can scale resources only when jobs run; because Milvus is purpose-built for large vector collections, the platform avoids the indexing and storage penalties of the previous search stack. The result is a cheaper, more stable recommendation architecture that matches analyst workflows and recommendation use cases.

Data flow

Periodic Serverless Spark jobs read product data from StarRocks, call a machine-learning model service to create title embeddings, batch-write vectors and metadata into Milvus, and then expose low-latency similarity search to recommendation and creator product-selection services.