Choose the best answer for each question. Answers are at the bottom.
Q1. Your generative AI app must respond within 200ms. Which endpoint type should you use? A) Batch endpoint B) Real-time managed online endpoint C) Serverless compute endpoint D) Scheduled pipeline
Q2. Which authentication mode is recommended for production endpoints? A) API key B) Anonymous access C) Azure AD (Entra ID) token D) Shared access signature
Q3. You want to test a new model with 5% of production traffic. What technique should you use? A) Blue-green deployment B) Canary deployment with traffic splitting C) Shadow deployment D) Rolling update
Q4. In a CI/CD pipeline, what should happen before deploying to production? A) Manual code review only B) Automated evaluation gate checking quality and safety C) Unit tests only D) Load testing only
Q5. Your groundedness score is 3.2 against a threshold of 4.0. What should happen? A) Deploy with a warning B) Block deployment and alert the team C) Deploy to staging only D) Retry evaluation
Q6. Which IaC tools can provision an AI Foundry hub? A) Bicep only B) Terraform only C) Both Bicep and Terraform D) Portal only
Q7. You need to process 100,000 documents overnight. Which is most cost-effective? A) Real-time endpoint with auto-scaling B) Batch endpoint C) Multiple parallel endpoints D) Serverless function
Q8. How should secrets be managed in CI/CD? A) Hardcoded in YAML B) GitHub Secrets or Azure Key Vault C) Command-line arguments D) Config files
Q9. Production error rate spikes to 15%. What is the fastest remediation? A) Redeploy previous model B) Shift traffic to last known good deployment C) Scale up D) Restart endpoint
Q10. What is the recommended environment promotion order? A) dev -> production B) staging -> dev -> production C) dev -> staging -> production D) dev -> QA -> staging -> production
Q11. Why register models with eval_score and dataset tags? A) Improves performance B) Enables traceability and reproducibility C) Required by Azure D) Speeds deployment
Q12. How should prompt templates be versioned? A) Database with timestamps B) Versioned files in source control C) Azure portal only D) No versioning needed
Q13. During canary deployment, P99 latency exceeds 5000ms at 20% traffic. What should happen? A) Continue to 50% B) Automatically roll back C) Scale up canary D) Wait 24 hours
Q14. What is the RPO for prompt configurations stored in Git? A) 15 minutes B) 1 hour C) Near zero D) 24 hours
Q15. Which GitHub auth method avoids storing long-lived secrets for Azure deployments? A) Personal access tokens B) Service principal with client secret C) OIDC with Workload Identity Federation D) SSH keys
| Q | Answer | Explanation |
|---|---|---|
| 1 | B | Real-time managed online endpoints provide low-latency synchronous inference. |
| 2 | C | Azure AD token auth is more secure with RBAC and no key rotation needed. |
| 3 | B | Canary deployment routes a small percentage of traffic to validate new versions. |
| 4 | B | Automated evaluation gates verify quality and safety before production. |
| 5 | B | Failing any threshold should block deployment to protect users. |
| 6 | C | Both Bicep and Terraform support Azure AI Foundry resources. |
| 7 | B | Batch endpoints are designed for high-volume async workloads at lower cost. |
| 8 | B | Secrets belong in secure stores like GitHub Secrets and Key Vault. |
| 9 | B | Traffic shifting is instant; redeployment takes minutes. |
| 10 | C | dev -> staging -> production is the standard promotion path. |
| 11 | B | Tags enable traceability for audits and reproducibility. |
| 12 | B | Source control provides history, diff, review, and rollback. |
| 13 | B | Automated rollback on metric degradation prevents spreading user impact. |
| 14 | C | Git commits are durable; prompt configs have near-zero RPO. |
| 15 | C | OIDC with Workload Identity Federation uses short-lived tokens. |