Microsoft AZ-104: VM Availability Options
Azure VM resilience starts with the failure a workload must tolerate. Availability sets, zones, and scale sets protect against different events, and none can compensate for an application with a single unprotected dependency. Before choosing one, identify whether the system must survive host maintenance, an entire facility outage, or demand growth—and whether the application can recover across that boundary.
On this page
- Identify the outage before choosing a VM feature
- Availability sets handle local infrastructure correlation
- Availability zones protect a broader local boundary
- Scale sets address the fleet rather than only placement
- Make state and network entry part of the decision
- Match maintenance strategy to the chosen construct
- Add regional recovery only when it answers a real requirement
- Choose deliberately and document what remains unprotected
Identify the outage before choosing a VM feature
Two identical virtual machines do not automatically make an application highly available. If both depend on one database, one private DNS resolver, or a load balancer that checks the wrong port, users can still experience a complete interruption. Availability design begins by drawing the request path and marking every component whose failure interrupts the business function. Azure offers several compute placement and scaling mechanisms, but their value depends on that inventory. A Microsoft AZ-104 availability decision should start with the failure being tolerated; zones, sets, and scale sets do not provide interchangeable protection.
Imagine an internal finance service that must remain usable during host maintenance but can accept a longer outage after a regional disaster. It may need distinct VM fault domains and a working load-balancing health probe without requiring a second Azure region. A public customer service with tighter objectives could justify separate availability zones and replicated state. The goal is not to use the most elaborate feature; it is to align the platform’s isolation boundary with the application’s recovery objective.
Capture recovery time and recovery point expectations separately. A compute configuration can allow another replica to serve traffic quickly while leaving stateful transactions vulnerable to data loss. Conversely, a well-protected database may survive but remain inaccessible if network or identity infrastructure is unavailable. Availability evidence must cover the complete workflow, not the VM resource status indicator.
Availability sets handle local infrastructure correlation
An availability set places related VMs into fault and update domains within the regional infrastructure. Fault domains are intended to limit shared hardware, power and network failures; update domains affect how planned platform maintenance is distributed. Microsoft documents both mechanisms and notes that availability sets do not isolate a workload from every datacenter-level event.
This can suit workloads operating in regions or designs where zone deployment is unavailable, and where independently managed VMs need protection from a single host-related interruption. The application still requires two functioning instances, traffic distribution, and a data architecture capable of handling failover. Putting one VM in an availability set does not create a second replica. Likewise, application patches that simultaneously break every instance remain an application-control problem.
Availability sets also have migration and lifecycle implications. A VM’s placement choice is not necessarily a harmless mutable property after deployment, so operators should review supported reprovisioning paths rather than assume it can be switched by editing a dropdown. Test how updates are applied, and record whether a dependency such as a managed disk or single shared appliance undermines the intended isolation.
Availability zones protect a broader local boundary
Availability zones are separate physical locations within supported Azure regions, with independently designed power, cooling and networking infrastructure. Deploying application replicas in different zones reduces the effect of a single-zone outage. This is a stronger isolation boundary than distributing VMs solely across an availability set, but it introduces connectivity, latency, and service-availability decisions.
Before pinning VMs to particular zones, verify that the desired size, disk configuration, networking service, and dependent platform services support the deployment in those locations. An application might span zones while its writable database, outbound firewall path, or regional gateway does not. Data replication mode must match the failure model, and the operator should understand whether a zonal interruption changes a customer-visible endpoint or simply shifts traffic to a healthy backend.
Zones are not a second region. A region-wide problem can still affect all its zones, and many Azure services have regional control-plane or operational dependencies. Distinguish a zonal design from a disaster-recovery plan; do not claim regional recovery because a diagram shows three isolated racks. Microsoft’s Azure VM availability options provide the supported placement families and related services.
Scale sets address the fleet rather than only placement
A Virtual Machine Scale Set manages multiple VM instances through a common control plane. That can simplify consistent configuration, instance replacement, autoscale decisions, and upgrades. Scale sets may participate in zone-based or regional placement depending on orchestration mode and settings; they are not inherently a substitute for thoughtful fault isolation.
Consider a stateless web tier experiencing strong weekday traffic peaks. A scale set can add instances based on approved metrics and remove capacity when demand falls. A standalone pair of VMs in an availability set cannot provide the same automated fleet lifecycle without additional management. But scaling blindly on CPU can also create failures: new instances might start before an application has warmed its caches, or a failed health probe may cause repeated replacement during a downstream database outage.
Microsoft currently recommends examining flexible orchestration for high availability when appropriate and explains zone and fault-domain options in its discussion of scale set reliability. Verify support for the actual VM and extension types, upgrade behavior, and zone capacity before treating that recommendation as a universal migration requirement for every legacy workload.
Make state and network entry part of the decision
Compute placement is only useful when clients can reach a healthy replica. A load balancer or application gateway must target the intended backend pool, check a meaningful health endpoint, and stop sending requests to an instance that cannot serve them. A web server returning a successful HTTP status while its required database is unavailable may fool an overly simple probe into keeping a broken instance in service.
For stateful applications, decide which component owns the authoritative data and how other instances observe it. Shared session state in one VM’s memory can make an apparently redundant pair behave as a single point of failure. Repeated writes to a single disk or file share require a storage design that supports the access pattern. Resilience may come from a managed data service, application replication, or recovery automation, but it does not arise from VM placement alone.
The operational design should also describe ingress and egress networking, private DNS, certificate handling, and update sequencing. If a node is drained for maintenance, can existing transactions finish? What happens to queued work? Do monitoring and health checks distinguish an application failure from transient infrastructure maintenance? These are questions of service continuity, not a checklist of Azure feature names.
Match maintenance strategy to the chosen construct
An update-domain strategy is not the same as an application deployment plan. Azure platform maintenance may be distributed across infrastructure boundaries, but a release pipeline that replaces every service instance at once can create its own outage. Plan rolling application upgrades or controlled deployment rings independently of placement settings.
Scale sets provide fleet-management controls, yet automatic upgrades still need safe health signals, rollback criteria, and enough remaining capacity. For independently managed VMs, a patch schedule should intentionally avoid taking all replicas offline at the same time. When zones are used, consider whether a planned configuration change could accidentally remove the only healthy instance in one or more zones.
A concrete maintenance rehearsal might disable a single test backend, verify that the probe removes it, update the backend, then return it to service. Record the user-observed outcome and the event timestamps. The result reveals whether the traffic-distribution layer and health check behave as documented; a successful Azure portal deployment alone does not prove rolling-service continuity.
Add regional recovery only when it answers a real requirement
If regional interruption is inside the risk assessment, availability sets and zones are insufficient by themselves. Some workloads need cross-region replicas, tested failover endpoints, secondary-region identity/network dependencies, or a deliberate Site Recovery workflow. Even with replicated VMs, recovering an application may require DNS, certificate, database, and secret access in the destination region.
Regional redundancy changes cost and operational complexity. Running a warm copy of an application involves continued resource charges and frequent configuration checks. A cold restoration strategy may be less expensive but slower. Compare approaches using actual tolerated interruption and acceptable data loss rather than assuming that a second region is always necessary or always too expensive.
Operators should rehearse failure scenarios at the level of the user transaction. A VM stopping is a useful injected fault, but an end-to-end test asks whether a client can complete the operation after the failure and whether persisted state is accurate. Restore the environment afterward and record which handoffs required manual intervention.
Choose deliberately and document what remains unprotected
A practical selection record lists the outage being handled, the replica count, the Azure placement or fleet-management construct, the health and routing design, the data-dependency strategy, and the recovery test that proves the arrangement. An availability set is reasonable for local hardware-related risks; zones serve stronger zonal isolation; scale sets support consistent fleets and elasticity. They can be combined in supported configurations, but none is a complete availability architecture by itself.
When an architecture diagram says “high availability,” ask which failure the deployment can survive, what evidence will show the surviving instance taking traffic, and whether the application remains useful afterward. Test a permitted failure in a non-production environment and record the actual outage observed. Zone placement, scale-set distribution, and a load-balancer probe address different parts of that outcome; no single label substitutes for testing the whole request path.