Design a Machine Learning Model Deployment Service (MLOps)

To design a robust MLOps system for Netflix, I would start by establishing a centralized Feature Store to ensure consistency between training and inference, preventing train-serving skew. For model management, I'd implement a strict versioning strategy using Git for code and DVC for data artifacts, allowing us to roll back instantly if a new model degrades performance. The core of the deployment strategy involves a Canary release pattern combined with Shadow Mode. In Shadow Mode, the candidate model processes live traffic alongside the production model but doesn't affect user recommendations; we collect these predictions to validate accuracy against ground truth data later. Once validated, we shift a small percentage of traffic to the new model. To handle data drift, I would deploy continuous monitoring using Evidently AI or custom statistical tests to detect shifts in user behavior patterns, such as sudden changes in viewing genres during holidays. If drift exceeds a threshold or A/B test metrics drop, an automated pipeline triggers an immediate rollback to the previous stable version, ensuring uninterrupted service for our global audience.