Kubernetes for ML Workloads Quiz

Answers and explanations appear as you take the quiz.

1. What is the primary job of the Kubernetes scheduler for ML workloads?
   • Writing the training code for data scientists
   • Deciding which node in the cluster runs each containerized pod
   • Labeling the training data automatically
   • Choosing the neural network architecture
2. How does Kubernetes typically make GPUs available to a pod?
   • Through device plugins that advertise resources like nvidia.com/gpu
   • By emailing the GPU vendor for a license
   • GPUs cannot be used in Kubernetes
   • By converting CPU cores into GPUs at runtime
3. With a NoSchedule taint on a GPU node, what does a matching pod toleration do?
   • Forces the pod onto that node regardless of capacity
   • Changes the model’s optimizer
   • Allows the pod to be considered for that node without guaranteeing placement
   • Encrypts every GPU tensor
4. Which tool adds ML-specific capabilities like distributed-training operators on top of Kubernetes?
   • Microsoft Word
   • Apache for serving static web pages
   • A spreadsheet macro
   • Kubeflow
5. Why is Kubernetes well suited to bursty ML workloads?
   • It schedules jobs onto free resources and frees that hardware when jobs finish
   • It permanently reserves all GPUs for one user
   • It prevents any job from ever scaling down
   • It only runs one container per cluster
6. What does 'scaling an inference endpoint to zero' mean on Kubernetes?
   • Deleting the model permanently
   • Setting accuracy to zero percent
   • Running zero copies when there's no traffic, then spinning up on the next request
   • Reducing the model's parameters to zero
7. One ML pod requests four GPUs, while two nodes each have only two available. Can ordinary scheduling combine those nodes for that one pod?
   • Yes; every pod automatically spans all cluster nodes
   • Yes; the scheduler converts CPU cores into the missing GPUs
   • Only if the pod name contains the word distributed
   • No; the pod must fit on one eligible node
8. An inference service scales to zero while idle. Which user-facing delay should its team test when traffic resumes?
   • The training-set labeling speed
   • Cold-start delay while a serving replica becomes ready
   • The file size of an unrelated notebook
   • The number of discarded vocabulary tokens