14.8.1: Review Questions

Review Questions

Multiple Choice

1. Which cloud service provides access to multiple foundation models (Claude, Llama, Stable Diffusion) via a single API without managing infrastructure?
   • a) AWS SageMaker
   • b) Amazon Bedrock
   • c) Google Kubernetes Engine
   • d) Azure DevOps
2. Which open-source framework enables running serverless workloads on Kubernetes across multiple cloud providers?
   • a) Docker
   • b) Terraform
   • c) Knative
   • d) Prometheus
3. Training a single large language model can consume energy equivalent to:
   • a) Charging a smartphone
   • b) Running a household for a week
   • c) The lifetime carbon emissions of several cars
   • d) Powering a small city for a year

Short Answer

4. Explain how serverless computing is evolving beyond simple stateless functions. Describe three specific advancements (e.g., stateful workflows, edge serverless, serverless containers) and their significance.
5. What is MLOps, and how does it relate to DevOps? Describe the key stages of an MLOps pipeline from data preparation to model monitoring. Why is MLOps more complex than traditional CI/CD?
6. Explain the concept of "Foundation Model as a Service." How do services like Amazon Bedrock and Azure OpenAI Service change who can build AI applications? What are the implications for cloud infrastructure and costs?
7. What is AIOps? Describe three specific ways AI is being used to manage cloud infrastructure operations. How might AIOps change the role of cloud administrators in the future?

True/False

8. True or False: AWS Lambda@Edge allows serverless functions to run at CDN edge locations closer to end users.
9. True or False: Distributed training always makes ML model training faster regardless of the model size or network bandwidth between nodes.
10. True or False: Carbon-aware scheduling means running workloads only during business hours to reduce energy consumption.

Answer Key:

1. b
2. c
3. c
4. (1) Stateful workflows: AWS Step Functions, Azure Durable Functions coordinate sequences of functions with preserved state. (2) Edge serverless: Lambda@Edge, Azure IoT Edge run functions at CDN/edge for low latency. (3) Serverless containers: Fargate, Cloud Run run arbitrary containers with auto-scaling and pay-per-use without managing servers.
5. MLOps = ML + DevOps: automated pipeline for ML lifecycle. Stages: data collection → data preparation → feature engineering → model training → model evaluation → model deployment → monitoring → retraining. More complex than CI/CD because: data versioning (not just code), model drift detection, experiment tracking, feature stores, GPU resource management.
6. FMaaS = cloud providers hosting pre-trained foundation models accessible via API. Amazon Bedrock provides Claude, Llama, etc. without hosting. Impact: democratizes AI (startups can build AI apps without GPU clusters), shifts cost from training to inference, creates new cloud spending category. Cloud providers become AI platforms.
7. AIOps = AI for IT Operations. Uses: (1) anomaly detection in metrics/logs (detecting unusual patterns before outages), (2) predictive scaling (forecasting demand spikes based on historical patterns), (3) automated remediation (auto-restarting failed services, auto-scaling based on ML predictions). Future: cloud admins shift from reactive firefighting to designing AI-driven self-healing systems.
8. True
9. False
10. False (it means scheduling workloads for times/regions with lower carbon intensity on the power grid)