The rush of AI workloads onto Kubernetes has quietly rewritten how architects pick AZ-305 compute solutions. Microsoft’s summer 2026 wave brought NCv6 GPU nodes to AKS, Anyscale-managed Ray in public preview, and automatic zone placement for node pools — and every one of those changes shows up in the design questions the exam now asks. Therefore, if your last mock test hinged on “VMs versus App Service,” expect this year’s paper to hinge on “which compute service best absorbs a bursty GPU inference workload without stranding capacity.”
This set continues our AZ-305 series. Previous installments covered cost optimization, identity and governance, data storage, business continuity, and a mixed-domain review. Specifically, Questions 51-60 focus on Domain 4 — Design infrastructure solutions — with an emphasis on compute choices: Virtual Machines, VM Scale Sets, App Service, Container Apps, AKS, Functions, and Batch. For example, several scenarios below force you to choose between two services that can both technically host the workload; only one meets the constraint the customer actually cares about.
As a result, treat each scenario as a design conversation, not a trivia question. Read the constraint first. In addition, watch for keywords like “scale to zero,” “regulated,” “burst,” “GPU,” “stateful,” and “hybrid” — those words are almost always the tiebreaker between the plausible answers.
Question 51: Event-Driven Microservices with Scale-to-Zero
Contoso is building a new order-processing platform composed of 30 small, stateless microservices that react to messages from Azure Service Bus. Traffic is spiky and often idle overnight. The team wants Linux containers, per-service autoscaling based on queue depth, and no charge when a service sits idle. Which compute service best fits these AZ-305 compute solutions requirements?
A) Azure Kubernetes Service with cluster autoscaler and KEDA
B) Azure Container Apps
C) Azure App Service on a Linux P1v3 plan
D) Azure Virtual Machine Scale Sets with Docker
👁 Reveal Answer
Correct Answer: B
Explanation: Azure Container Apps is a serverless Kubernetes-backed PaaS with built-in KEDA that scales each app independently — including down to zero — based on HTTP traffic or event sources like Service Bus. AKS (A) also uses KEDA, but you still pay for at least one node in the pool and take on cluster upkeep, which contradicts the “no charge when idle” and “minimal ops” implication. App Service plans (C) cannot scale to zero — you pay for the plan continuously. VMSS with Docker (D) requires you to build the scaling and orchestration yourself and never scales to true zero.
Question 52: Regulated Workload Isolation
Woodgrove Bank must run a set of internal .NET web APIs that handle regulated payment data. Compliance requires the compute plane to be fully isolated from other tenants, reachable only over private endpoints, and deployed into the bank’s own virtual network. The team already uses App Service elsewhere and wants to keep that programming model. Which design meets these AZ-305 compute solutions constraints with the least redesign?
A) Deploy the APIs to a multitenant App Service Premium v3 plan with VNet integration
B) Deploy the APIs to App Service Environment v3 in the bank’s VNet
C) Rehost the APIs on Azure Container Apps with a Dedicated workload profile
D) Rehost the APIs on Azure Kubernetes Service with private cluster enabled
👁 Reveal Answer
Correct Answer: B
Explanation: App Service Environment v3 (ASEv3) provides a single-tenant, dedicated App Service deployment inside your VNet — exactly what “fully isolated” and “keep the App Service programming model” require. Multitenant Premium v3 with VNet integration (A) still shares underlying compute with other customers, so it fails the isolation requirement. Container Apps (C) and AKS (D) both meet the isolation and networking asks technically, but each forces a rehost onto a container platform, contradicting the “least redesign” constraint.
Question 53: Serverless Functions with VNet and Cold-Start Sensitivity
Fabrikam runs an inventory reconciliation function that must reach a database behind a private endpoint. The workload is idle 20 hours a day, but during business hours it must respond within 400 ms without cold-start delays. Cost-wise, leadership refuses to pay for always-hot Premium capacity around the clock. Which AZ-305 compute solutions option best meets these requirements?
A) Azure Functions on the legacy Consumption plan
B) Azure Functions on the Premium (Elastic Premium) plan
C) Azure Functions on the Flex Consumption plan with a small number of always-ready instances
D) Azure Functions in a Dedicated App Service plan (S1)
👁 Reveal Answer
Correct Answer: C
Explanation: Flex Consumption combines serverless economics (scale to zero when idle) with VNet integration and a small pool of always-ready instances that eliminate cold starts during known busy windows. The legacy Consumption plan (A) does not offer VNet integration in the classic model and suffers cold starts. Premium (B) meets the technical asks but forces you to pay for at least one warm instance 24/7, which leadership rejected. A Dedicated App Service plan (D) also runs 24/7 and, at S1, would not scale as elastically. Therefore Flex Consumption is the tightest fit.
Question 54: VMSS Orchestration Mode Choice
Northwind Traders is designing a stateless batch worker fleet that must span three availability zones, use both Spot and Regular VMs in one pool, and support instance-level operations such as attaching an existing NIC to a specific VM. Which VM Scale Set orchestration mode should the architect choose?
A) Uniform orchestration mode
B) Flexible orchestration mode
C) Availability Set with autoscale
D) Azure Batch pool in User Subscription mode
👁 Reveal Answer
Correct Answer: B
Explanation: Flexible orchestration mode is the current recommended mode for VM Scale Sets and, unlike Uniform, supports mixing VM sizes and priorities (Spot + Regular) in the same scale set and treats each VM as an individually addressable resource. Uniform (A) treats all instances as identical and does not mix Spot and Regular in one pool. Availability Sets (C) provide fault-domain protection but no zonal spread and no native autoscale. Azure Batch (D) is a great HPC service but is overkill for a “stateless batch worker fleet” that just needs mixed-priority VMs across zones.
Question 55: GPU Inference for a Trending AI Workload
Tailwind Traders wants to host a real-time recommendation model that must serve 50 requests per second with sub-second latency, uses NVIDIA GPUs, and is packaged as a container. Data science already runs experiments in Azure Machine Learning, but production inference must live near the retail microservices already on AKS. What is the best design?
A) Deploy the model as an Azure Machine Learning managed online endpoint on GPU compute
B) Add a GPU-enabled node pool (NCv6-series) to the existing AKS cluster and deploy the model as a pod
C) Rehost as an Azure Function in the Flex Consumption plan
D) Run the container on Azure Container Instances with a burstable SKU
👁 Reveal Answer
Correct Answer: B
Explanation: Because production inference “must live near the retail microservices already on AKS,” adding a dedicated GPU node pool (for example the NCv6-series) to the existing cluster keeps latency low and reuses the platform the team already operates. Azure ML online endpoints (A) work well, but they add a separate control plane, extra hop, and duplicated ops for a workload that is otherwise on AKS. Functions (C) do not offer GPUs. Azure Container Instances (D) supports GPUs in limited SKUs and is not designed for sustained high-throughput inference at 50 rps.
Question 56: Legacy Windows .NET Framework Application
Litware, Inc. must lift an on-premises ASP.NET (full .NET Framework 4.8) application to Azure with as little code change as possible. The app depends on IIS-specific modules and a Windows-only third-party DLL. Which target minimizes rework?
A) Refactor the application to .NET 8 and deploy to Azure Container Apps
B) Package as a Windows container and deploy to Azure Kubernetes Service
C) Deploy to Azure App Service on a Windows plan
D) Rebuild as an Azure Function on the Flex Consumption plan
👁 Reveal Answer
Correct Answer: C
Explanation: Azure App Service on a Windows plan natively runs classic ASP.NET on IIS with essentially no code change, which directly satisfies “as little code change as possible.” Refactoring to .NET 8 (A) is a rewrite. Windows containers on AKS (B) work but demand containerization, node-pool management, and operational lift for a single legacy app. Rebuilding as a Function (D) requires re-architecting the entire application. Therefore App Service Windows is the least-friction target.
Question 57: HPC-Style Embarrassingly Parallel Batch
The engineering team at Adventure Works needs to run 200,000 short, independent Monte Carlo simulations each night. Jobs are packaged as command-line executables, take between 30 seconds and 5 minutes each, and must complete within a 4-hour window. Cost per run matters more than latency. Which service is the best fit?
A) Azure Batch with a pool of Spot VMs and low-priority queue
B) Azure Functions on Flex Consumption
C) AKS with a large node pool and a Job controller
D) Azure Container Apps jobs on the Consumption plan
👁 Reveal Answer
Correct Answer: A
Explanation: Azure Batch was purpose-built for large, embarrassingly parallel HPC workloads with a job scheduler, per-task retry, and native support for Spot (low-priority) VMs — exactly the “200,000 short executables, cost matters most” pattern. Functions (B) can run parallel jobs but hit orchestration limits and are more expensive per compute-second at this scale. AKS (C) works but requires you to build the scheduling and Spot-node handling yourself. Container Apps jobs (D) are viable for smaller batches, but Batch remains the exam’s canonical answer for HPC-scale, cost-sensitive parallel compute.
Question 58: Availability SLA for a Single-Region Web Tier
Your team is designing a two-VM web tier in a single region for a customer that requires a 99.99% VM SLA. Cost is a secondary concern; the customer accepts either zonal or non-zonal deployment options. Which configuration meets the SLA target?
A) Deploy both VMs in a single Availability Set with two fault domains
B) Deploy both VMs in a single availability zone, in a Proximity Placement Group
C) Deploy the two VMs across two availability zones in the region
D) Deploy both VMs to a single zone with Premium SSD v2 disks
👁 Reveal Answer
Correct Answer: C
Explanation: Two or more VMs spread across two or more availability zones qualify for Azure’s 99.99% VM SLA. An Availability Set (A) tops out at 99.95% for two or more VMs. A single-zone deployment (B or D) is a single failure domain and only qualifies for the 99.9% single-instance Premium SSD SLA at best. As a result, only zonal spread meets the requirement.
Question 59: Spot vs Reserved for a Stateless Rendering Farm
Fabrikam Studios runs a stateless video-rendering farm that can tolerate interruptions and retries. Rendering workloads run 24/7 with a stable baseline of about 30 nodes and unpredictable bursts up to 300 nodes. Finance wants the lowest steady-state cost without leaving too much on the table when bursts hit. Which purchase pattern fits?
A) 300 nodes on Pay-as-you-go
B) 300 nodes on 3-year Reserved Instances
C) 30 nodes on 3-year Reserved Instances, bursts on Spot VMs
D) 30 nodes on Spot, bursts on Reserved Instances
👁 Reveal Answer
Correct Answer: C
Explanation: Reserved Instances (or Savings Plans) give the deepest discount on predictable steady-state capacity — the 30-node baseline. Spot VMs give the deepest discount on interruptible capacity — perfect for the burst-up-to-300 tail, since the workload is stateless and tolerates eviction. Buying reservations for the peak (B) wastes money most of the day. All Pay-as-you-go (A) leaves the reservation discount on the table. Reserving the burst but leaving the baseline on Spot (D) is backwards and risks losing steady capacity to evictions.
Question 60: Multi-Cluster Kubernetes Governance
Contoso operates 14 AKS clusters across four Azure regions and two on-premises data centers. The platform team needs one place to enforce policy, distribute workloads, and manage cluster lifecycle for both the Azure and on-premises clusters. Which combination best meets the requirement?
A) Azure Kubernetes Fleet Manager only
B) Azure Arc-enabled Kubernetes only
C) Azure Kubernetes Fleet Manager for the AKS clusters and Azure Arc-enabled Kubernetes to project the on-premises clusters into Azure
D) Azure Policy assigned to each subscription independently
👁 Reveal Answer
Correct Answer: C
Explanation: Fleet Manager provides multi-cluster orchestration (workload propagation, update runs, load balancing) but only for AKS. Arc-enabled Kubernetes brings the on-premises clusters into the Azure control plane so they can receive Azure Policy, GitOps configurations, and monitoring. Fleet alone (A) cannot reach on-premises clusters. Arc alone (B) does not offer Fleet’s cross-cluster workload orchestration. Per-subscription Azure Policy (D) fragments governance and cannot manage lifecycle. Therefore the combination is the correct AZ-305 compute solutions answer.
Study Tips for These AZ-305 Compute Solutions
Use the following checklist as you drill this set of AZ-305 compute solutions — it maps to the eliminators the exam actually rewards.
- Anchor on the constraint, not the service. The exam rarely asks “what is X”; it asks “which service satisfies this specific constraint.” Underline the constraint (scale-to-zero, isolated, GPU, cold-start, cost) before you look at the answers.
- Memorize the SLA table. Single VM with Premium SSD = 99.9%, Availability Set = 99.95%, Availability Zones (2+) = 99.99%. This one table decides three or four questions on every real paper.
- Know the container ladder. App Service → Container Apps → AKS moves from “least ops, least control” to “most ops, most control.” Container Apps is Kubernetes without the Kubernetes — reach for it when the workload is event-driven, Linux, and needs to scale to zero.
- Flex Consumption is the new default for serverless Functions. It combines scale-to-zero, VNet integration, and always-ready instances. However, keep the Premium plan on your radar for workloads that require durable warm capacity beyond a small always-ready pool.
- Spot for interruptible, Reserved for predictable. If a workload is stateless and tolerates eviction, Spot almost always wins. For steady baseline, Reserved Instances or Savings Plans beat everything else.
Keep Practicing These AZ-305 Compute Solutions
Compute design is the single largest domain on the exam, so nailing this AZ-305 compute solutions set will move your score more than almost any other topic. Come back tomorrow for Questions 61-70, which will move into networking design — hub-and-spoke, Virtual WAN, and Private Link patterns.
Want more practice? Continue with our earlier sets: the Identity, Governance, and Monitoring drill (Questions 11-20), the Business Continuity set (Questions 31-40), and the Mixed Domain Review (Questions 41-50) to keep the whole exam blueprint fresh in your head.
For the official blueprint, review the Microsoft Learn Design infrastructure solutions learning path and cross-check any service defaults you are unsure about against current Microsoft documentation before exam day.
