AI-200 Containerized AI Workloads — 10 Questions

Question 1

You are configuring an Azure Container Apps environment with a custom Virtual Network. Your AI containers need to securely connect to an Azure SQL Database that is secured behind a private endpoint. Which configuration is required on the Container Apps environment?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Deploy the Container Apps environment into a custom VNet with workload profile subnets. To reach an Azure SQL Database behind a private endpoint, the Container Apps environment must be deployed into a custom VNet (VNet-injected) so its workloads get private IPs in your address space and can resolve the privatelink.database.windows.net DNS record. Workload profile subnets provide the dedicated compute infrastructure required for this VNet-integrated deployment. This is the only option that establishes the required network path.

Answer

Use an external Azure Application Gateway in front of the container apps

Answer

Enable Azure Bastion on the subnet

Answer

Configure IP restriction rules on the ingress controller

Question 2

You are deploying a Python Flask API that wraps an ONNX runtime model to Azure Container Instances (ACI). The container needs to expose port 5000 externally. Which parameter must you specify when creating the container group?

Accepted Answer

--dns-name-label and port 5000 mapping. To expose port 5000 externally on an Azure Container Instance, you must specify a DNS name label (which gives the container group a public FQDN) and map the port in the container group's IP address configuration. The --dns-name-label and --ports parameters together create the public endpoint that routes to the container's listening port.

Answer

--assign-identity

Answer

--restart-policy Never

Answer

--os-type Windows

Question 3

You have an AI model container that takes 10 seconds to load weights into memory upon startup. Users report HTTP 503 errors immediately after deployment because traffic is routed to the container before it is ready. Which probe should you configure?

Accepted Answer

Startup probe. A startup probe is specifically designed to handle slow-starting containers by delaying the application of liveness and readiness checks until the container has fully initialized. In this scenario, the model takes 10 seconds to load weights, so a startup probe will prevent premature traffic routing and avoid HTTP 503 errors. Once the startup probe succeeds, the container is considered started, and normal liveness/readiness probes take over.

Answer

Liveness probe

Answer

Resource quota probe

Answer

TCP socket probe

Question 4

You are managing an AKS cluster and need to ensure that GPU driver updates do not disrupt running AI inference workloads. Which AKS feature helps manage GPU operator lifecycle and driver installation?

Accepted Answer

NVIDIA GPU Operator installed via Helm. The NVIDIA GPU Operator, deployed via Helm, is the standard mechanism for managing the full lifecycle of GPU software components on Kubernetes — including driver installation, container runtime hooks, device plugin, and DCGM monitoring. On AKS GPU node pools, installing the operator via Helm allows driver updates to be orchestrated as Kubernetes-managed DaemonSets, so updates can be rolled out in a controlled fashion without disrupting running inference workloads. This is the AKS-recommended approach for managing GPU driver lifecycle beyond the default pre-installed drivers.

Answer

Azure Bastion host extension

Answer

Azure Policy for Kubernetes

Answer

Container Insights agent upgrade

Question 5

You are scaling an AI workload in Azure Container Apps. Which TWO scale rule triggers are natively supported without custom KEDA extensions?

Accepted Answer

HTTP concurrent requests. Azure Container Apps natively supports HTTP-based scaling rules, so option A (HTTP concurrent requests) is correct because the platform exposes an HTTP scale trigger that scales replicas based on the number of concurrent HTTP requests hitting the container app's ingress endpoint. Option E (CPU and Memory utilization thresholds) is also correct because Container Apps provides built-in CPU and memory utilization scale triggers that monitor the average resource usage of replicas and scale accordingly, without requiring any custom KEDA scaler. Options B, C, and D are incorrect because Azure VM disk I/O throughput, Active Directory login failure rate, and hardware temperature sensors are not native Container Apps scale triggers; they would require external metrics or custom KEDA extensions to be used as scaling criteria.

Answer

Azure Virtual Machine disk I/O throughput

Answer

Active directory login failure rate

Answer

Custom hardware temperature sensors

Question 6

You are optimizing resource allocation for a GPU-backed AI model deployment on AKS. Which THREE Kubernetes concepts or objects are essential for managing GPU resources effectively?

Accepted Answer

Resource requests and limits specifying nvidia.com/gpu. Managing GPUs in AKS requires node selectors, resource limits for nvidia.com/gpu, and tolerations for GPU taints.

Answer

Node taints and tolerations for GPU node pools

Answer

Horizontal Pod Autoscaler based on Azure Load Balancer SNAT ports

Answer

Kubernetes persistent volume claims for GPU driver binaries

Question 7

Which TWO Azure services can store and serve private Docker container images for AI workloads?

Accepted Answer

Azure Container Instances (pulling from ACR). Azure Container Registry (D) is Microsoft's private, managed Docker registry service, purpose-built to store and serve private container images (including AI workload images) via the Docker Registry HTTP API v2, with authentication through Microsoft Entra ID, service principals, or admin credentials. Azure Container Instances (C) is correct because ACI can pull and run images directly from a private ACR using registry credentials or a managed identity, thereby serving those private container images to run AI workloads without managing a VM host. Azure Cognitive Services (A) provides prebuilt AI APIs and models, not a container image registry, so it cannot store or serve private Docker images. Azure Bastion (B) is a managed jump-host service for secure RDP/SSH access to VMs and has no image registry capability. Azure Logic Apps (E) is a workflow automation service and does not store or serve container images.

Answer

Azure Cognitive Services

Answer

Azure Bastion

Answer

Azure Logic Apps

Question 8

You are configuring an Azure Kubernetes Service (AKS) cluster to host GPU-accelerated AI model inference workloads using NVIDIA GPUs. Which component must you install in the cluster to enable containerized workloads to utilize the GPU nodes?

Accepted Answer

NVIDIA GPU Operator. The NVIDIA GPU Operator is the supported way to provision GPU drivers, the device plugin, container runtime hooks, and monitoring on AKS GPU node pools. It bundles the device plugin and driver lifecycle so containers can request nvidia.com/gpu resources.

Answer

NVIDIA Device Plugin for Kubernetes

Answer

KEDA scaler

Answer

Azure Monitor agent

Question 9

You manage an AKS cluster running high-throughput AI inference pods. During traffic spikes, pods experience latency due to slow container image pulls from Azure Container Registry (ACR). You want to optimize image pull performance across cluster nodes. Which feature should you enable?

Accepted Answer

AKS Artifact Streaming (formerly Image Streaming). AKS supports Artifact Streaming (previously called Image Streaming), which streams container image layers from ACR on demand instead of fully pulling the image before the container starts. This reduces pod startup latency for large AI inference images by fetching layers lazily over the network as the container accesses them, so pods can begin running before the entire image is downloaded. This directly addresses slow image pulls during traffic spikes.

Answer

ACR Tasks auto-purge

Answer

ACR Geo-Replication

Answer

Azure Blob Storage CDN integration

Question 10

You are designing a high-performance AI inference architecture on Azure Kubernetes Service (AKS). Which THREE AKS features or configurations should you incorporate to optimize performance and resource utilization? (Choose three.)

Accepted Answer

Cluster Autoscaler to automatically scale node pool size based on pending pod demands. Option A is correct because the Cluster Autoscaler dynamically adjusts the number of nodes in a node pool based on pending pod resource requests, ensuring that AI inference workloads get the capacity they need without manual intervention and improving overall resource utilization. Option C is correct because AI inference typically requires GPU acceleration, and AKS supports GPU-optimized node pools using NC or ND series VMs, which provide the necessary compute for model inference. Option E is correct because node taints and tolerations allow you to dedicate GPU nodes exclusively to AI inference pods, preventing non-GPU workloads from consuming these expensive resources and ensuring predictable performance. Option B is incorrect because single-core, low-memory burstable B-series VMs are not suitable for high-performance AI inference, which demands significant CPU, memory, and GPU resources. Option D is incorrect because disabling the Kubernetes scheduler and manually placing every container instance would remove the benefits of automated scheduling, reduce efficiency, and is not a recommended practice for optimizing performance in AKS.

Answer

Running all pods exclusively on single-core, low-memory burstable VMs (B-series)

Answer

Disabling the Kubernetes scheduler to manually place every container instance