Shadow deployment runs a new model alongside production without affecting users, validating predictions in real conditions. How it works: Production traffic duplicated to shadow model. Shadow predictions logged but not served. Compare predictions to production model and ground truth. Purpose: Validate new model on real traffic patterns before promoting. Catch issues without user impact. What to evaluate: Prediction agreement with production, latency performance, resource usage, edge case handling, error rates. Comparison methods: Log both predictions, compare offline. Statistical analysis of disagreements. Manual review of interesting cases. Duration: Run until confident - typically days to weeks depending on traffic volume and variability. Infrastructure needs: Duplicate inference pipeline, logging infrastructure, comparison tooling, minimal latency impact on production. When to use: High-risk model changes, new architectures, major retraining, regulated environments. Progression: Shadow mode, then canary deployment, then full rollout. Risk mitigation. Limitations: Does not catch issues from actually serving predictions (feedback loops, user behavior changes).