Design a production-ready machine learning system for real-time personalized recommendations serving 10 million users. Your architecture should address: (1) data pipeline design for processing billions of daily events, (2) model training infrastructure with continuous retraining, (3) serving infrastructure for sub-100ms latency, (4) monitoring for model performance and data drift, (5) A/B testing framework for model evaluation, and (6) cold-start handling for new users and items. Explain your choices for each component and how they scale as the system grows.