Model retraining periodically updates model weights on fresh data to maintain performance as distributions shift. Why retrain: Combat data drift and concept drift, incorporate new patterns, improve on mistakes, adapt to changing world. Retraining strategies: Scheduled: Fixed intervals (daily, weekly, monthly). Simple but may miss urgent needs. Triggered: When performance degrades below threshold or drift detected. Responsive but complex. Continuous: Online learning with streaming data. Always current but harder to manage. What to keep: Architecture, hyperparameters (unless tuning), training pipeline. What changes: Training data (add recent, possibly remove old), weights. Data windows: Use all historical data, sliding window (last N months), weighted by recency, or combination. Validation: Always validate new model before deployment. A/B test or shadow mode. Automation: Automated retraining pipelines detect trigger, retrain, validate, deploy. Full MLOps. Challenges: Training compute costs, validation time, rollback planning, handling concept drift mid-training. Best practice: Monitor continuously, retrain proactively, validate thoroughly before promotion.