Better search relevance using Learning to Rank at mobile.de [PyCon DE & PyData Berlin 2024]

Python was chosen as the primary language for building the search ranking service due to team familiarity, ecosystem maturity, and existing data science workflows

The service architecture uses a two-stage approach:

• Stage 1: Candidate retrieval using Elasticsearch/Java to find top 1000 matches
• Stage 2: Re-ranking using Python service to get final top 20 results

MongoDB was selected over Redis for caching because it offers:

• Native vector embedding support
• Fast aggregation pipelines
• Easy business logic integration
• Better support for complex queries

Key technical implementation details:

• FastAPI used as web framework for async support and documentation
• Docker containers for deployment
• Kubernetes for orchestration
• Position bias handled through feature engineering
• XGBoost 2.0 used for ranking model

Critical requirements defined upfront:

• 30ms latency requirement
• Support for business logic/boosting
• Pagination capability
• High availability
• Stateless design

Testing and deployment best practices:

• Comprehensive unit tests
• Local testing with Docker Compose
• A/B testing capabilities
• Small Docker image sizes
• Automated CI/CD pipeline

Key learnings:

• Start with clear requirements documentation
• Consider long-term maintainability
• Handle feature availability carefully
• Use cache-based approach for performance
• Monitor business metrics closely

The project improved unique conversion metrics and enabled data scientists to deploy models independently without backend engineering bottlenecks

Service achieved 99.3% successful request rate with ~6ms response times

Two-sided marketplace considerations required balancing seller visibility with buyer relevance through multi-objective optimization