CAIC Cloud Infrastructure 3 — Questions and Answers
Question 1: What does 'infrastructure as code' (IaC) provide for AI platform teams?
- Automated model training
- Reproducible, version-controlled environment provisioning (Correct answer)
- Real-time inference optimization
- Automatic hyperparameter search
Correct answer: Reproducible, version-controlled environment provisioning
IaC tools like Terraform and Pulumi let teams define cloud resources in code, enabling repeatable, auditable, and diff-able infrastructure changes.
Question 2: A CAIC consultant recommends separating AI training and inference workloads into different node pools. What is the PRIMARY reason?
- Training and inference require different hardware profiles and scaling behaviors (Correct answer)
- Cloud providers charge less for separate pools
- Inference needs more CPU cores than training
- It simplifies billing by combining costs
Correct answer: Training and inference require different hardware profiles and scaling behaviors
Training is batch-oriented and GPU-heavy, while inference is latency-sensitive and may need different GPU types or even CPU-only nodes, requiring distinct pools.
Question 3: Which networking concept reduces AI data transfer costs when moving large datasets between cloud storage and compute within the same region?
- Content Delivery Network (CDN)
- Private endpoints / VPC peering (Correct answer)
- Public internet transit
- WAN optimization appliance
Correct answer: Private endpoints / VPC peering
Private endpoints and VPC peering keep traffic on the provider's backbone, avoiding public internet egress fees and reducing latency.
Question 4: What is the purpose of a feature store in cloud AI infrastructure?
- To store raw unprocessed log files
- To centralize, share, and serve ML features consistently across training and serving (Correct answer)
- To manage container image versions
- To provide GPU driver updates
Correct answer: To centralize, share, and serve ML features consistently across training and serving
A feature store (e.g., Feast, Tecton) ensures the same feature logic is used in training and inference, preventing training-serving skew.
Question 5: An organization needs to run AI workloads that require physical hardware control and data residency, but also want elastic capacity. Which pattern fits best?
- All-in public cloud
- On-premises only
- Hybrid cloud with cloud bursting (Correct answer)
- Serverless-only architecture
Correct answer: Hybrid cloud with cloud bursting
Hybrid cloud with cloud bursting keeps sensitive workloads on-premises while automatically routing overflow to the public cloud during peak demand.
Question 6: Which metric is most important when evaluating cloud GPU instance performance for distributed AI training?
- Disk IOPS
- Inter-GPU bandwidth (e.g., NVLink or InfiniBand throughput) (Correct answer)
- Public IP address allocation speed
- DNS resolution latency
Correct answer: Inter-GPU bandwidth (e.g., NVLink or InfiniBand throughput)
Distributed training relies on frequent gradient synchronization between GPUs; high inter-GPU bandwidth (NVLink, InfiniBand) directly reduces communication bottlenecks.
Question 7: What is the role of a service mesh (e.g., Istio) in a cloud-native AI inference platform?
- Compiling ML models to native code
- Managing secure, observable, and policy-driven service-to-service communication (Correct answer)
- Storing trained model artifacts
- Scheduling GPU jobs
Correct answer: Managing secure, observable, and policy-driven service-to-service communication
A service mesh provides mTLS encryption, traffic policies, retries, and distributed tracing between microservices including inference endpoints.
CAIC Cloud Infrastructure 3
What does 'infrastructure as code' (IaC) provide for AI platform teams?