Home/ Part XIII — Expert Mode: Systems, Agents, and Automation/42. Fine-Tuning vs Prompting vs Retrieval (Decision Framework)/42.5 Deployment and rollback strategy

42.5 Deployment and rollback strategy

Overview and links for this section of the guide.

On this page

Deployment

// Gradual rollout
const modelRouter = {
  route(request: Request) {
    // Start with 5% traffic
    if (random() < 0.05) {
      return fineTunedModel;
    }
    return baseModel;
  }
};

// Monitor both models
trackMetrics('base', baseModel);
trackMetrics('fine-tuned', fineTunedModel);

// If fine-tuned performs better, increase traffic

Rollback

// Instant rollback capability
const ACTIVE_MODEL = process.env.ACTIVE_MODEL || 'base';

// If issues detected, change env var and restart
// No redeployment needed

// Always keep base model available
// Never delete working model versions

Where to go next