High Availability Design
AZ-305 AZ-305 · Design business continuity solutions
High availability is the art of designing so that failure is an event, not an outage. Azure provides the building blocks — availability zones, availability sets, scale sets, and zone-redundant services — and the architect's job is to combine them into a design that meets the availability target without paying for resilience the business does not need.
Availability zones and availability sets
An availability zone is a physically separate datacentre (or group of datacentres) within an Azure region, with independent power, cooling, and networking. Deploying across zones protects against datacentre-level failure. A zonal deployment pins resources to specific zones; a zone-redundant deployment spreads them automatically across zones. Availability sets are the older mechanism for VMs within a single datacentre: VMs are spread across fault domains (racks sharing power/network — protecting against hardware failure) and update domains (groups rebooted together during platform maintenance — protecting against maintenance downtime). The exam's key distinction: availability sets protect within a datacentre; availability zones protect against datacentre failure. For new designs, zones are the default — availability sets remain relevant only in regions without zones, or for VM types that do not support zones. Note the SLA differences: two VMs in an availability set carry a 99.95 percent SLA; two VMs across zones carry a 99.99 percent SLA; a single VM with premium storage carries 99.9 percent.
Scaling for availability: Virtual Machine Scale Sets
Virtual Machine Scale Sets (VMSS) deploy and manage a set of identical, load-balanced VMs with autoscaling. In flexible orchestration mode, VMSS supports zone-spanning deployments with fault-domain spreading, combining the benefits of zones and sets. Autoscale rules add instances when CPU or queue length rises and remove them when load falls, so availability and cost scale together. For stateless workloads, VMSS with zone redundancy and a load balancer is the standard HA pattern. Stateful workloads need shared storage (Azure Files, managed disks with shared access) or a data-tier HA design so any instance can serve any request.
Zone-redundant PaaS services
Many PaaS services offer zone redundancy as a configuration choice: Azure SQL Database (Business Critical, General Purpose, Hyperscale), App Service (Premium v3 and above in supported regions), Azure Kubernetes Service (system node pools across zones), Azure Cache for Redis (Premium/Enterprise), Service Bus (Premium), Event Hubs (Standard and above), and Cosmos DB (multi-region writes with zone redundancy within each region). The design rule is to enable zone redundancy for every stateful PaaS service in a production workload — it is usually a checkbox with a modest cost premium. For services without zone redundancy, plan the failover: a regional deployment with Traffic Manager or Front Door can route around a regional failure, but the data tier must be replicated (failover groups, geo-replication) or the failover is useless.
Active-active vs active-passive
Active-passive (hot standby) keeps the secondary region idle or scaled down until failover — lower cost, but failover takes time and the standby must be tested. Active-active runs both regions live behind a global load balancer (Front Door or Traffic Manager), giving the fastest recovery and the best global latency — but it demands a data tier that supports multi-region writes or conflict-free replication, which is the hard part. Cosmos DB multi-region writes and Azure SQL active geo-replication support active-active patterns; most relational workloads are simpler as active-passive with failover groups. The exam tests this trade-off: active-active is not automatically better — it is better only when the RTO and the data tier justify its complexity and cost.
SLAs and composite availability
Every Azure service publishes an SLA, and the composite SLA of a solution is the product of its components' SLAs: two services at 99.9 percent each give 99.8 percent combined. This is why the exam includes SLA arithmetic: if the business requires 99.95 percent, a chain of 99.9 percent services will not deliver it, and the design must add redundancy. Also remember that SLAs are targets backed by service credits, not guarantees, and that the SLA applies only when the documented configuration is followed (for example, two VMs in an availability set, not one). Design for the SLA the business needs, document the composite, and avoid single points of failure — every tier (compute, data, network) needs its own redundancy story.
Health probes and graceful degradation
Availability is not just redundancy — it is detection and response. Health probes on load balancers and Application Gateway detect unhealthy instances and stop sending them traffic; without probes, a load balancer happily sends requests to a dead server. Design health endpoints that check dependencies (database connectivity, not just "the process is running"). Plan graceful degradation: if a non-critical dependency fails, the application should serve cached or reduced content rather than failing entirely. And use circuit breakers and retry with backoff so transient failures do not cascade — patterns the exam rewards in application architecture questions too.
Designing for failure: beyond redundancy
Redundancy is necessary but not sufficient for high availability. The application must also be designed to fail gracefully: use retry policies with exponential backoff for transient failures, circuit breakers to stop calling a failing dependency and fail fast instead, bulkheads to isolate failures so one dependency cannot exhaust all resources, and timeouts on every external call so a hung dependency does not hang the application. Chaos engineering — deliberately injecting failures in test environments — validates that these patterns actually work; Azure Chaos Studio provides managed fault injection for VMs, AKS, and other services. For multi-region data, the hardest problem is conflict resolution: if two regions accept writes to the same record, which write wins? Cosmos DB offers last-writer-wins and custom conflict resolution; relational databases generally avoid the problem by keeping one writable region. When an exam scenario describes a multi-region active-active design, check the data tier first — if it cannot handle multi-region writes, the design cannot be active-active however good the load balancing is.
Measuring availability
Availability targets are meaningless without measurement. Azure Monitor availability tests (URL ping and multi-step web tests) probe your application from multiple locations and record uptime; Application Insights availability reports the percentage of successful tests over time. Define your service-level objectives (SLOs) from the business requirement (for example, 99.95 percent availability measured monthly) and build error budgets from the gap between the SLO and 100 percent — the budget is the amount of downtime you can "spend" on deployments and experiments. When the budget is exhausted, freeze risky changes until reliability recovers. Alert on burn rate (how fast the error budget is being consumed) rather than on single failures, so the team is paged for trends that threaten the SLO, not for every transient blip. This SLO-and-error-budget approach is straight from the Well-Architected Framework's reliability pillar and appears in exam scenarios about operations.
Exam traps to watch for
- Using availability sets when availability zones are available — zones protect against datacentre failure; sets only protect within a datacentre.
- Forgetting the data tier in an active-active design — compute can fail over in seconds, but it is useless without replicated data.
- Assuming a single VM is highly available — a single VM's SLA is 99.9 percent at best; HA needs at least two instances with a load balancer.
- Multiplying SLAs incorrectly — the composite SLA is the product of component SLAs, so every added dependency lowers total availability.
Practice questions
Test yourself on this guide. Answers are checked instantly, with a full explanation for each question.
Q1.A company runs two VMs in a single datacentre and needs to protect against a datacentre power failure. The region supports availability zones. What should the architect recommend?
Explanation: Availability zones are physically separate datacentres, so spreading VMs across zones with a zone-redundant load balancer survives a datacentre power failure. Option B is wrong because availability sets spread VMs within one datacentre only. Option D is wrong because backup restores data; it does not keep the service running during an outage. Option C is wrong because more VMs in the same datacentre does not help when the datacentre loses power.
Exam objective: AZ-305 Domain 3: Design for high availability
Q2.A web application must maintain 99.95 percent availability. The design uses a single App Service instance and a single Azure SQL Database. What is the main problem with this design?
Explanation: Single instances have no redundancy: if either the App Service instance or the database fails, the application is down, and the composite SLA of two single-instance services falls below 99.95 percent. Option C is wrong because App Service with multiple instances across zones can exceed 99.95 percent. Option D is wrong because Azure SQL Database has built-in HA with zone redundancy available. Option A is wrong because a CDN improves content delivery performance, not application-tier availability.
Exam objective: AZ-305 Domain 3: Design for high availability
Q3.A stateless web application experiences traffic spikes that triple normal load for a few hours each day. The company wants automatic scaling and zone redundancy. What should the architect recommend?
Explanation: VMSS with autoscale adds instances during spikes and removes them afterwards, zone deployment protects against datacentre failure, and the load balancer distributes traffic — all requirements met. Option B is wrong because a single VM cannot scale or survive zone failure. Option C is wrong because a fixed availability set cannot autoscale and protects only within one datacentre. Option D is wrong because Site Recovery is disaster recovery, not load-based scaling.
Exam objective: AZ-305 Domain 3: Design for high availability
Q4.A solution chains three services, each with a 99.9 percent SLA. What is the composite SLA of the solution, and what does it imply?
Explanation: The composite SLA is the product: 0.999 × 0.999 × 0.999 = 0.997, or 99.7 percent — below a 99.9 percent target, so the design needs redundancy (or higher-SLA components) to meet it. Option B is wrong because composite SLA is multiplicative, not the minimum. Option D is wrong because 99.97 percent is the wrong arithmetic (0.999³ ≈ 0.997). Option C is wrong because adding dependencies lowers availability, it never raises it.
Exam objective: AZ-305 Domain 3: Design for high availability
Q5.A global application must serve users from the nearest region with automatic failover, and the data tier must support writes in multiple regions. Which combination should the architect recommend?
Explanation: Front Door provides global anycast routing to the nearest healthy region, and Cosmos DB multi-region writes let each region accept writes — the active-active pattern the scenario requires. Option B is wrong because SQL failover groups are active-passive: only one region accepts writes. Option C is wrong because Load Balancer is regional, not global. Option D is wrong because Application Gateway is regional and single-region Cosmos writes cannot serve multi-region write traffic.
Exam objective: AZ-305 Domain 3: Design for high availability
Q6.A load balancer continues to send traffic to a web server whose application has crashed, although the VM is still running. What should the architect recommend to fix this?
Explanation: A health probe that exercises the application (including its dependencies) detects the crashed application and removes the instance from the load balancer pool. Option A is wrong because more VMs do not stop traffic reaching the broken one. Option D is wrong because autoscale responds to load metrics, not application health. Option B is wrong because availability sets address hardware and maintenance failure, not application-level health detection.
Exam objective: AZ-305 Domain 3: Design for high availability
Q7.A company deploys an AKS cluster for a production workload and needs the cluster to survive a datacentre failure. What should the architect recommend?
Explanation: Spreading the AKS system node pool across availability zones keeps the Kubernetes control plane and nodes available when a datacentre fails. Option B is wrong because availability sets protect only within a datacentre. Option D is wrong because backup restores data; it does not keep the cluster running. Option C is wrong because a single-zone deployment fails with its zone, regardless of autoscale.
Exam objective: AZ-305 Domain 3: Design for high availability
Q8.A retail website's recommendation service occasionally fails, and when it does, the entire product page fails to load. What design pattern should the architect recommend so the page still loads without recommendations?
Explanation: Graceful degradation lets the page render without the non-critical recommendation feature when its service fails, preserving the core user experience. Option C is wrong because active-active reduces failure probability but does not change what happens when the feature is unavailable. Option D is wrong because a higher tier does not guarantee the service never fails. Option B is wrong because a second load balancer adds redundancy to traffic distribution, not to the application's failure behaviour.
Exam objective: AZ-305 Domain 3: Design for high availability
Q9.A company needs 99.99 percent availability for a pair of VMs. Which configuration delivers that SLA?
Explanation: Two VMs deployed across availability zones carry a 99.99 percent SLA. Option C is wrong because two VMs in an availability set carry a 99.95 percent SLA. Option D is wrong because a single VM with premium storage carries a 99.9 percent SLA. Option A is wrong because two VMs in the same fault domain share a failure domain and have no HA SLA.
Exam objective: AZ-305 Domain 3: Design for high availability