Trendyol Speeds Up Search Ranking Model Deployment

Trendyol's Search Relevance team re-engineered how ranking models reach production for e-commerce search. Previously, each new model meant translating a data scientist's spreadsheet into a dedicated Go file often exceeding a thousand lines, then testing, compiling and deploying it manually — a two-to-three-day cycle per model that could consume most of a sprint when multiple A/B challengers were involved.

To cut this 'model lead time,' the team moved to a config-driven architecture. Data scientists now write a compact 150-200 line YAML 'model signature' describing which features the model needs, from where, and in what order. This file passes through a validation pipeline that automatically opens a merge request in the new ranking service, leaving engineers to simply review and approve.

The new service parses this three-part YAML (model info, features, inference) to handle feature fetching, dependency resolution and inference-input assembly on its own, without hardcoded logic or recompilation. The result is a platform that preserves the same ranking request flow while dramatically shortening iteration time and removing the manual handoff between data science and engineering.