AI/ML Workloads Quiz
Review allocation, training/serving, storage, networking, security, observability and cost boundaries.
Quiz Questions
1. How should a Pod request GPUs exposed by the NVIDIA device plugin?
- A. Set only requests with nvidia.com/gpu:0.5
- B. Specify integer nvidia.com/gpu limits, with an equal request if also supplied
- C. Add only an nvidia.com/gpu node annotation
- D. Name a RuntimeClass nvidia-mps to enable sharing automatically
Show Answer
Answer: B. Specify integer nvidia.com/gpu limits, with an equal request if also supplied
Limits alone imply an equal request. Allocatable GPU capacity is not a node label. GPU Feature Discovery labels, MIG, MPS and time-slicing require separate configuration; a shared slot does not guarantee a fraction of GPU memory.
2. Which installed legacy Kubeflow resource understands TensorFlow distributed roles?
- A. Deployment automatically creates all TF_CONFIG values
- B. Job guarantees training success and simultaneous worker placement
- C. TFJob provides framework-specific roles/configuration
- D. TFJob is built into Kubernetes and needs no installation
Show Answer
Answer: C. TFJob provides framework-specific roles/configuration
Install its CRD/controller and configure restart policy, framework strategy, checkpoints, storage and networking. Recovery and gang scheduling are not universal guarantees. A regular Job can also train when actual execution code is supplied.
3. How should an existing FSx for Lustre filesystem be connected through CSI?
- A. Combine a Lustre CR and a StorageClass containing the existing ID
- B. Use a static PV with volumeHandle/dnsname/mountname and a bound PVC
- C. SCRATCH_2 supports every persistent backup/throughput option
- D. FIO against unmounted /data measures FSx
Show Answer
Answer: B. Use a static PV with volumeHandle/dnsname/mountname and a bound PVC
Static attachment differs from dynamic filesystem creation. Compare capacity, IOPS/throughput, latency, access patterns, permissions, AZ/network and backups. Benchmark a dedicated test path on the actual mounted volume.
4. What should guide AI/ML resource allocation?
- A. Every training job requires a GPU
- B. Measured CPU/memory/GPU use, model size, I/O, latency and throughput
- C. Always request maximum resources
- D. Omitting requests is always most efficient
Show Answer
Answer: B. Measured CPU/memory/GPU use, model size, I/O, latency and throughput
CPU and other accelerators can be appropriate. OMP/MKL thread settings affect supporting libraries, not GPU sharing. CUDA allocator settings differ from Kubernetes GPU allocation/isolation. VPA effects depend on mode, restart behavior and workload compatibility.
5. What should be checked when selecting a serving platform such as KServe?
- A. Any model URI immediately works
- B. Runtime image, model format, protocol, storage and deployment mode match requirements
- C. A normal Deployment cannot serve inference
- D. All modes offer identical canary and scale-to-zero behavior
Show Answer
Answer: B. Runtime image, model format, protocol, storage and deployment mode match requirements
A Deployment can run an inference server. KServe Knative/Standard and versioned Seldon APIs differ. Explainability, processing, weighted routing and batch inference are not universally enabled defaults. TorchServe is not a maintained default for new use.
6. What is required for HPA scaling on GPU utilization or request load?
- A. Allocated nvidia.com/gpu automatically becomes a utilization Resource metric
- B. Exporters and custom/external metrics adapters with correct labels, units and aggregation
- C. Increase replicas to fix low model accuracy
- D. Attach several HPAs to the same Deployment for additive scaling
Show Answer
Answer: B. Exporters and custom/external metrics adapters with correct labels, units and aggregation
The standard metrics-server Resource path exposes CPU/memory. GPU/request metrics need separate collection/adapters. Pod metrics must map to namespace/Pod labels; a global average can lose that mapping. Use one scaling owner and validate whether load/queue is a better proportional signal than latency percentiles.
7. Which statement about distributed-training networking is correct?
- A. A zone annotation alone controls placement
- B. GPU hostPath mounts configure EFA/GPUDirect
- C. Validate node-label placement, hardware, drivers/plugins, communication libraries and network policies together
- D. NCCL always transports data through MPI
Show Answer
Answer: C. Validate node-label placement, hardware, drivers/plugins, communication libraries and network policies together
NCCL over EFA uses AWS OFI NCCL and libfabric; MPI can launch processes. Multus/SR-IOV, MTU and RDMA configuration are environment-specific. Generic NIC examples are not turnkey EKS settings. NetworkPolicy can affect availability/performance by blocking required traffic.
8. Which model/data protection statement is correct?
- A. Secret base64 is encryption
- B. Kubernetes RBAC grants S3/KMS permissions
- C. Configure API RBAC, workload IAM, encryption/key management, file credentials and networking separately
- D. Store large models in Secrets and decryption keys in environment variables by default
Show Answer
Answer: C. Configure API RBAC, workload IAM, encryption/key management, file credentials and networking separately
Secrets have size limits and base64 is encoding. Use object storage with authorization/integrity checks for large models. A Pod service account does not always need Secret get permission to mount a Secret volume: API reads and kubelet volume delivery differ. PodSecurityPolicy was removed; use current admission/PSS controls.
9. What should ML service observability distinguish?
- A. GPU utilization determines accuracy
- B. ServiceMonitor selects Pod labels directly
- C. Service errors/latency/throughput, resource use and ground-truth model quality are separate
- D. All container logs are Docker JSON
Show Answer
Answer: C. Service errors/latency/throughput, resource use and ground-truth model quality are separate
ServiceMonitor selects Service labels/named ports. Define whether latency includes failures and avoid counting one request as both success and error. Quality needs labels/ground truth. Validate CRI framing versus app JSON, Grafana datasource/panel schema, labeled alert aggregation and zero-traffic behavior.
10. How should cost optimization be validated?
- A. Use Spot for every workload
- B. Newest instances are always cheapest
- C. Measure required performance, total cost, interruption recovery/checkpointing and actual Pod/node reclamation
- D. Nighttime automatically lowers On-Demand rates
Show Answer
Answer: C. Measure required performance, total cost, interruption recovery/checkpointing and actual Pod/node reclamation
Fewer Pods do not imply EC2 termination. Check availability, quotas, AMI/drivers and NodePool limits. Training taints need matching tolerations; mixed CPU/GPU groups are not EKS Hybrid Nodes. Creating an unused Autoscaler ConfigMap does not change controller flags.