AI Model Wiki

    Skip to main content
    20 min
    Lesson 3 of 4

    Safe AI Deployments

    From the course AI CI/CD Pipelines

    Implement safe deployment strategies for AI applications with rollback capabilities.

    Deployment Strategies

    Canary Deployments

    # Kubernetes canary config
    apiVersion: argoproj.io/v1alpha1
    kind: Rollout
    metadata:
      name: ai-service
    spec:
      strategy:
        canary:
          steps:
            - setWeight: 5
            - pause: {duration: 10m}
            - analysis:
                templates:
                  - templateName: ai-quality-check
            - setWeight: 25
            - pause: {duration: 10m}
            - setWeight: 50
            - pause: {duration: 10m}
            - setWeight: 100
    

    Feature Flags for AI

    from launchdarkly import LDClient
    
    ld_client = LDClient("sdk-key")
    
    def get_ai_response(user, prompt):
        # Check which model version to use
        model_version = ld_client.variation(
            "ai-model-version",
            user,
            default="gpt-4"
        )
        
        # Check if new prompt template is enabled
        use_new_prompt = ld_client.variation(
            "new-prompt-template",
            user,
            default=False
        )
        
        return call_model(
            model=model_version,
            prompt=new_prompt if use_new_prompt else legacy_prompt
        )
    

    Rollback Automation

    def automated_rollback_check():
        metrics = get_production_metrics(window="5m")
        
        rollback_triggers = [
            metrics["error_rate"] > 0.1,
            metrics["latency_p99"] > 5000,
            metrics["user_feedback_negative"] > 0.3,
            metrics["hallucination_rate"] > 0.05,
        ]
        
        if any(rollback_triggers):
            trigger_rollback()
            alert_oncall("AI deployment rolled back due to quality issues")
    

    Key Takeaways

    • •Canary deployments gradually shift traffic (5% → 25% → 50% → 100%) with analysis steps
    • •Feature flags enable instant rollback of model versions and prompt templates
    • •Automated rollback triggers on error rate, latency, feedback, and hallucination spikes
    • •Argo Rollouts integrates AI quality checks into Kubernetes deployment pipelines
    0 of 4 lessons marked complete

    Progress is kept in this browser only.