Knowledge Check - Containers and AI Platform
Test your understanding of the key concepts from Module 6. Select the best answer for each question, then submit to see your results. You need to score at least 60% to pass.
FinCorp moves its portal and partner API services into the public Managed Kubernetes cluster fc-k8s-portal. The services need host and path routing, TLS termination and health checks at one internet entry point, the same one that already fronts the portal VMs. A developer proposes a Service of type LoadBalancer for each microservice. What should the architect recommend?
A LoadBalancer service does not provision an external load balancer; it attaches a static IP to one worker node, with up to 2 Gbit/s per node, so one service per microservice multiplies single-node entry points without Layer 7 routing. Manifests do not create managed load balancers, so the ALB is built separately and routes to an in-cluster ingress controller. Private node pools do not support LoadBalancer services at all.
A nightly batch job in fc-k8s-ai needs eight nodes for two hours. The finance team asks for the lowest cost outside that window and proposes an autoscaled node pool with a minimum of zero nodes. How should the architect design the batch capacity?
The cluster autoscaler cannot reduce a node pool to zero, so every autoscaled pool keeps at least one billable node. One shared batch pool keeps the idle cost to a single warm node, while a pool per job multiplies that floor. The control plane is free of charge, but it does not suspend the node pools, which are billable Compute Engine servers.
FinCorp's security officer requires that the AI workers' nodes accept no connections from the internet and that only the administration network can use the Kubernetes API. The team plans a cluster with private node pools and assumes the API is then private too. What completes the design?
Private node pools isolate the data plane: nodes have internal addresses only and reach the internet outbound through the NAT Gateway the service deploys on the reserved address. The control plane remains exposed to the internet and is protected by restricting IP access. Node pool nodes cannot be added to Network Security Groups, and public pools would expose the nodes, contradicting the first requirement.
Three CI/CD pipelines push images to FinCorp's Private Container Registry, and two clusters pull them. The audit team wants a leaked credential to reach as little as possible and to be stoppable at once. Which access design should the architect choose?
Registry access is defined by token scopes, so least privilege means one narrowly scoped token per consumer, with Push only where a pipeline publishes. A token's state can be disabled at once, while expiry deletes it, so rotation before the expiry date is planned. A wildcard Admin token gives every leak full reach, and repositories cannot be made public, since authentication is always required.
A FinCorp business unit wants Red Hat OpenShift instead of Managed Kubernetes. The compliance office asks who runs the control plane and who provides the evidence for its operation. What should the architect answer?
IONOS CLOUD does not provide a managed OpenShift: Red Hat CCSP partners deploy and manage the clusters after a Red Hat validation, and Rancher Prime is self-managed by the customer. The infrastructure is the same, but the operation of a customer-deployed control plane is a workload whose evidence the partner or the customer must produce. Managed Kubernetes is the option where IONOS CLOUD runs the control plane.
FinCorp builds an advisor assistant that answers from internal policy documents, and later a credit pre-assessment on the same stack. Customer data must stay in Germany and must not train any model. Which design meets the requirements?
AI Model Hub processes inference in Germany and does not use customer data for training, and RAG brings the documents to the model without training it. As deployer, FinCorp pseudonymizes request content and, for a high-risk use such as credit, builds its own logging and human oversight, because no pre-configured AI Act logging template is provided. The SLA does not apply to Cloud GPU VMs, and fine-tuning on customer records would contradict the requirement that customer data trains no model.