Skip to content

AZ-900 Azure architecture, compute and networking

The second AZ-900 skill area, Describe Azure architecture and services, is the largest on the exam at 35-40% (skills outline as of July 20, 2026). This page covers its first two objectives: the core architectural components of Azure, and Azure compute and networking services. Storage and identity are on the next page.

The running example is OrdersPortal, an invented order-management app (ordersportal.example.test) with a web front end, an API and a SQL database, which its team is deploying to Azure.

The Azure CLI fragments on this page were not run here (no Azure subscription is available to this site’s build); their syntax was checked against the Azure CLI reference and Microsoft Learn samples. Run them in Azure Cloud Shell or a local Azure CLI after az login.

Topic What the exam expects
Regions and availability zones Regions, region pairs, sovereign regions, availability zones, datacenters
Resource hierarchy Resources, resource groups, subscriptions, management groups and how they nest
Virtual machines VMs, scale sets, availability sets, Azure Virtual Desktop, VM resources
Containers and Functions Container Instances, Container Apps, AKS, Azure Functions; comparing compute types
Application hosting Web apps, containers and VMs as hosting options
Virtual networking Virtual networks, subnets, peering, public and private endpoints
Hybrid connectivity VPN Gateway and ExpressRoute
Azure DNS Hosting DNS zones in Azure
Load balancing and traffic routing Load Balancer, Application Gateway, Front Door, Traffic Manager (context beyond the outline)

Azure’s physical infrastructure starts with datacenters: facilities with racks of servers, each with its own power, cooling and networking. You never pick a datacenter directly; you pick a region.

A region is a set of datacenters in a geographic area, connected by a low-latency network. Azure has more than 70 regions. Regions are grouped into geographies (for example, the United States or Europe), and each geography is a data residency boundary: choose regions in the right geography when data must stay in a country or market. Some services are global and have no region.

Availability zones are physically separate groups of one or more datacenters within a region, each with independent power, cooling and networking. Regions that support zones have at least three. Zones are usually within 100 kilometers of each other: close enough for low-latency replication, far enough apart that a local outage is unlikely to hit two at once. Services use zones in two ways: zonal resources are pinned to one zone you choose (for example, a virtual machine), and zone-redundant resources are spread across zones by the platform (for example, zone-redundant storage).

Region pairs: Microsoft pairs some regions with another region, usually in the same geography. Some services use the pair for geo-replication (geo-redundant storage replicates to the paired region), and planned platform updates are rolled out to one region of a pair at a time. You can’t choose the pairs, many newer regions are not paired and rely on availability zones instead, and you can still deploy a secondary copy to any region you like.

Sovereign regions are isolated instances of Azure for compliance or legal reasons, such as Azure Government for US government agencies and Azure in China, operated by 21Vianet. They often have a smaller set of services.

Terminal window
# List the regions available to your subscription (fragment, not run here)
az account list-locations --query "[].{name:name, displayName:displayName}" --output table

Pitfalls

  • A region pair does not give you automatic failover for your app. You still design and test disaster recovery.
  • Availability zone “1” in one subscription isn’t necessarily the same physical zone as “1” in another subscription.

A resource is anything you create in Azure: a virtual machine, a virtual network, a database. Every resource belongs to exactly one resource group, a logical container for resources that share a lifecycle. Resource groups can’t be nested or renamed, a resource can be moved between groups for many resource types, and deleting a resource group deletes everything in it. Access and policies applied to a group flow down to its resources.

A subscription is a unit of management, billing and scale. It links to an account in Microsoft Entra ID and is both a billing boundary (each subscription gets its own invoice and cost reports) and an access control boundary (access policies apply per subscription). Teams commonly create separate subscriptions for environments (development, test, production), for organizational units, or for billing.

Management groups sit above subscriptions. Policies and role assignments applied to a management group are inherited by every subscription, resource group and resource below it. Management groups can be nested up to six levels deep (not counting the root level or the subscription level), every Microsoft Entra tenant has one Tenant Root Group at the top, a directory supports up to 10,000 management groups, and each management group or subscription has exactly one parent.

The hierarchy from top to bottom: management groups > subscriptions > resource groups > resources.

For OrdersPortal, the team creates sub-ordersportal-prod and sub-ordersportal-dev under a Commerce management group, with one resource group per environment and workload:

Terminal window
# Create a resource group with tags for cost reporting (fragment, not run here)
az group create \
--name rg-ordersportal-prod \
--location eastus2 \
--tags app=OrdersPortal env=prod costCenter=CC-1001

The command prints the new resource group as JSON, including "provisioningState": "Succeeded".

Pitfalls

  • Deleting a resource group is the fastest way to remove a test environment and also the fastest way to lose a production one. Use resource locks.
  • A resource group’s location only stores its metadata; the resources inside it can be in other regions.

Azure Virtual Machines are IaaS: you get a virtualized server with full control over the operating system and installed software, and you remain responsible for patching and configuring it. Typical uses are test and development, lift-and-shift migration, datacenter extension and disaster recovery.

Creating a VM provisions several resources that the exam expects you to know: the VM size (CPU cores, memory and purpose), managed disks for the operating system and data, a network interface, a virtual network and subnet, and optionally a public IP address and network security group. Size names encode the family: in Standard_D2s_v5, D is the general-purpose family, 2 the vCPU count, s support for Premium SSD and v5 the hardware generation.

  • Virtual Machine Scale Sets create and manage a group of identical, load-balanced VMs and can scale the count out or in automatically, on metrics or on a schedule.
  • Availability sets spread VMs across fault domains (servers that share power and network) and update domains (groups rebooted together during planned maintenance), so one hardware failure or maintenance event doesn’t take every VM down. The set itself is free; you pay for the VMs. In regions with availability zones, zone-based designs give broader isolation.
  • Azure Virtual Desktop is a desktop and app virtualization service: users connect from any device to Windows desktops and apps hosted in Azure, with Microsoft Entra ID for identity. It supports single-session and multi-session Windows.

Pitfall: deleting a VM doesn’t necessarily delete its disks, network interface or public IP address. Orphaned resources keep billing.

Containers package an application with its dependencies and share the host’s operating system kernel, so they start faster and are lighter than VMs. You don’t manage an operating system per container. Azure offers three main ways to run them:

  • Azure Container Instances (ACI): the fastest, simplest way to run a container without managing VMs (PaaS).
  • Azure Container Apps: also PaaS, with built-in load balancing and scaling, aimed at microservices.
  • Azure Kubernetes Service (AKS): managed Kubernetes for orchestrating many containers.

Azure Functions is event-driven, serverless compute. An event (HTTP request, timer, queue message) wakes the function, it scales automatically with demand, and on consumption-based hosting you pay only while code runs. Functions are stateless by default; Durable Functions keep state.

Need Choose
Full OS control, custom software, existing server image Virtual machine
Package an app and run it quickly or as microservices Containers
Short, event-driven pieces of code with variable demand Functions

Pitfall: containers and Functions still run on servers. “No servers to manage” is the benefit, not “no infrastructure”.

To host a web app or API you can use VMs (most control, most management), containers (portable, can scale per component) or Azure App Service. App Service is a PaaS offering for web apps, API apps, WebJobs (background tasks) and mobile back ends, on Windows or Linux, in languages including .NET, Java, Node.js, Python and PHP. It handles the infrastructure, load balancing, autoscaling and patching, and integrates with GitHub and Azure DevOps for continuous deployment.

For OrdersPortal, App Service is the natural fit for the web front end and API: the team deploys code from GitHub Actions and never logs in to a server. A static front end could also use Azure Static Web Apps.

Pitfall: App Service runs your code, but you still own app configuration, secrets and scaling rules.

An Azure virtual network (VNet) lets Azure resources communicate with each other, the internet and on-premises networks. Its capabilities, as listed on Microsoft Learn: isolation and segmentation, internet communications, communication between Azure resources, communication with on-premises resources, routing and filtering network traffic, and connecting virtual networks.

You define a private address space in CIDR notation and divide it into subnets. For example, 10.0.0.0/16 contains 65,536 addresses and can hold subnets such as 10.0.0.0/24 for the web tier and 10.0.1.0/24 for the data tier. Network security groups filter traffic with allow and deny rules on source, destination, port and protocol; network virtual appliances such as firewalls add deeper inspection. Virtual network peering connects two VNets, even in different regions, over the Microsoft backbone so traffic never crosses the public internet.

Public endpoints have a public IP address and are reachable from anywhere on the internet. Private endpoints sit inside a VNet with a private IP address from its address space, so you can reach a service such as a storage account or SQL database without exposing it publicly.

Terminal window
# VNet with a web subnet, then a data subnet (fragment, not run here)
az network vnet create \
--resource-group rg-ordersportal-prod \
--name vnet-ordersportal-prod \
--address-prefixes 10.0.0.0/16 \
--subnet-name snet-web \
--subnet-prefixes 10.0.0.0/24
az network vnet subnet create \
--resource-group rg-ordersportal-prod \
--vnet-name vnet-ordersportal-prod \
--name snet-data \
--address-prefixes 10.0.1.0/24

Pitfalls

  • Peered VNets must not have overlapping address spaces. Plan ranges before you need to connect networks.
  • Peering is not transitive by default: if A peers with B and B with C, A can’t reach C without extra configuration.

Azure VPN Gateway is a virtual network gateway that sends encrypted traffic over the public internet. It supports site-to-site connections (an on-premises VPN device to a VNet), point-to-site connections (an individual computer to a VNet) and network-to-network connections (VNet to VNet). A VNet can have only one VPN gateway, but one gateway can connect to several sites. Gateways are policy-based or route-based; route-based is required for point-to-site, VNet-to-VNet, multisite and coexistence with ExpressRoute. By default a gateway runs as two instances in active/standby; active/active and zone-redundant gateways add resilience.

Azure ExpressRoute extends an on-premises network into the Microsoft cloud over a private connection through a connectivity provider, called an ExpressRoute circuit. Traffic doesn’t travel over the public internet, which gives more reliability, faster speeds, consistent latency and higher security. It reaches Azure services and Microsoft 365, uses BGP for dynamic routing, has built-in redundancy at each peering location, and ExpressRoute Global Reach can connect your own on-premises sites to each other through the Microsoft network.

VPN Gateway ExpressRoute
Path Encrypted tunnel over the internet Private connection through a provider
Bandwidth and latency Limited by internet path Higher, more consistent
Typical use Branch offices, remote workers, quick setup Datacenter integration, large or regulated data transfer

Pitfall: even with ExpressRoute, DNS queries, certificate revocation checks and Azure Content Delivery Network requests still go over the public internet. A common resilient design uses a VPN gateway as the failover path for an ExpressRoute circuit.

Azure DNS hosts DNS domains on Azure’s global network of name servers, using anycast so the nearest server answers each query. You manage records with the same credentials, RBAC, activity logs, resource locks and billing as other Azure resources. It supports private DNS zones for custom names inside virtual networks, and alias records that point at an Azure resource (a public IP address, a Traffic Manager profile or a CDN endpoint) and update automatically when the resource’s IP address changes.

OrdersPortal buys ordersportal.example.test from a registrar, delegates it to Azure DNS, and manages its www and api records next to the rest of its infrastructure.

Pitfall: you can’t buy a domain name with Azure DNS. Buy it from a registrar (or as an App Service domain), then host the zone in Azure DNS.

Load balancing is not a named item in the current skills outline, but these services appear across the AZ-900 learning path and in architecture questions. Azure classifies them on two axes: global or regional, and HTTP(S) or non-HTTP(S).

Service Scope Layer Use it for
Azure Load Balancer Regional (also global) Layer 4, TCP and UDP Distributing traffic across VMs or scale set instances, public or internal
Application Gateway Regional Layer 7, HTTP(S) TLS termination, URL path and host routing, session affinity, optional Web Application Firewall
Azure Front Door Global Layer 7, HTTP(S) Global web apps: edge caching, acceleration, WAF, multiregion failover
Traffic Manager Global DNS-based Directing clients to endpoints in different regions or outside Azure

Traffic Manager works at the DNS level: it answers a DNS query with the best endpoint and the client then connects directly, so it never proxies traffic. Every profile monitors endpoint health and fails over automatically, and supports six routing methods: Priority (primary with backups), Weighted (split by weight, for example for a gradual rollout), Performance (lowest-latency endpoint), Geographic (by the location of the DNS query, for data sovereignty), MultiValue (return several healthy IPv4 or IPv6 endpoints) and Subnet (map client IP ranges to endpoints). Because clients cache DNS answers, failover speed depends on the record’s time to live.

Many designs combine services: Front Door globally in front of a regional Application Gateway, with an internal Load Balancer in front of non-HTTP back ends.

Pitfall: Load Balancer and Traffic Manager can carry HTTP traffic but provide no Layer 7 features such as path-based routing or WAF.

  1. What is the smallest number of availability zones in a region that supports them? Three. Some regions have more.

  2. You apply an Azure Policy to a management group. Which resources does it affect? Every subscription, resource group and resource beneath that management group, because policies and role assignments are inherited down the hierarchy.

  3. Which service hosts a web app without you managing virtual machines, with built-in autoscaling and deployment from GitHub? Azure App Service (PaaS).

  4. A company needs a private connection to Azure that does not traverse the public internet. Which service? Azure ExpressRoute. A VPN gateway encrypts traffic but still sends it over the internet.

  5. Can you register ordersportal.example.test in Azure DNS? No. Azure DNS hosts zones and records; you buy the domain from a registrar and delegate it to Azure DNS.

  6. Scenario: OrdersPortal must survive a datacenter failure in its main region, keep the database off the public internet and serve users in Europe and North America with low latency. Outline the design. Deploy the web tier across availability zones (zone-redundant App Service or a scale set spread over zones) so one datacenter failure doesn’t stop it. Put the database behind a private endpoint in the VNet and deny public network access. Add a second region in the other continent and route users with Azure Front Door (HTTP) or Traffic Manager Performance routing, with health probes for automatic failover. Replicate data to the second region and test the failover, since neither zones nor region pairs provide disaster recovery for the app automatically.