Shadow Deployment
Running a new model in parallel with production, processing real traffic without serving its predictions.
Overview
Shadow deployment (shadow mode) involves deploying a new ML model alongside the production model, where both process identical live traffic but only the existing model's predictions are served to users. The shadow model's predictions are logged and compared against the production model for evaluation.
Key Details
This approach provides realistic performance assessment without any user impact or risk. It helps catch issues that offline evaluation misses, such as latency problems, edge cases in production data, and integration issues. Shadow deployments are particularly valuable for high-stakes applications (finance, healthcare) where incorrect predictions could cause significant harm.