Microsoft AZ-104: Azure Well-Architected Operations
Azure architecture quality cannot be reduced to a security score or uptime target. Operationally successful workloads balance reliability, security, cost, performance, and the ability to change them safely. A design review is incomplete if the application meets its immediate functional need but cannot be observed, maintained, or recovered. Translate reliability, security, cost, and operational-excellence principles into concrete administrative checks that can be repeated after deployment.
On this page
- Establish what the workload must achieve
- Define reliability as a tested service outcome
- Make security boundaries understandable
- Optimize cost from usage patterns, not slogans
- Test performance under representative load
- Make operational excellence observable in production
- Examine trade-offs as specific decisions
- Turn architecture review into an operational experiment
Establish what the workload must achieve
Architecture discussions become unproductive when teams choose services before agreeing on the business function, the users affected by failure, and the consequences of data loss. A public checkout service, an internal monthly reporting pipeline, and a long-term archive do not need identical designs. Their reliability, security, performance, cost and operational requirements differ even when they use the same Azure subscription. A Microsoft AZ-104 design review should connect resilience, security, cost, and operations to the actual workload instead of applying every best practice mechanically.
Microsoft’s Well-Architected Framework presents five connected pillars: Reliability, Security, Cost Optimization, Operational Excellence and Performance Efficiency. They are lenses for explaining trade-offs, not five independent grades that can be maximized without consequence. A workload can improve one concern while making another more difficult or expensive.
Write a short workload charter before judging the platform: who owns the service, what transaction matters, what interruption is tolerable, which data must remain protected, what peak demand looks like, and how it will be supported. The charter turns abstract architectural advice into testable design decisions. Without it, a team may buy costly redundancy for a noncritical batch process while leaving a critical identity dependency unmanaged.
Define reliability as a tested service outcome
Reliable infrastructure is not merely a collection of redundant resources. Users must still complete their intended operation after a supported failure. Multiple application replicas may be ineffective if they share a single-zone database, unavailable secret store or misconfigured load-balancer health probe. Review the full dependency path and distinguish component availability from end-to-end service continuity.
An application that promises restoration within an hour needs evidence that recovery can be completed in that time. Examine backup and replication capabilities, routing, identity services, service configuration, escalation and manual approvals. A well-documented backup and recovery strategy may be more relevant than another live compute replica for a workload whose dominant risk is accidental data deletion.
Reliability has cost and complexity trade-offs. A secondary region increases resilience to some disasters, but adds networking, deployment, secret management and testing responsibilities. If nobody maintains the second region, the nominal protection may not exist when needed. Accept the simplest design that credibly meets documented recovery objectives and record where it does not.
Make security boundaries understandable
Security starts with clear identities, data classifications and permitted operations. A deployment pipeline should not hold permanent subscription Owner merely because that was convenient for an early test. A storage account handling sensitive records should have a deliberate identity and network model rather than defaulting to wide access and promising to add restrictions later.
Use separate controls for separate purposes. Azure RBAC governs resource and data operations at supported scopes, Policy evaluates configurations, locks restrict selected management changes, and network controls govern packet paths. They can complement one another, but none is universal protection. Azure RBAC scope deserves the same architectural attention as the subnet diagram because misassigned privileges can undermine otherwise careful isolation.
Security also affects operations. Extremely restrictive changes that no one has tested can break routine recovery, while emergency workarounds may silently bypass the original standard. The security design should include controlled elevation, log access, documented exceptions and a tested way to restore a critical service without exposing unrelated systems.
Optimize cost from usage patterns, not slogans
The cost pillar concerns the value of resources as much as their price. Choosing the lowest-cost SKU may fail if it cannot meet demand or recover within the service’s objectives. Overprovisioning every environment for worst-case production traffic is equally hard to defend. Compare the real utilization, growth pattern and business value of the capacity purchased.
Consider a report-processing job that runs for two hours each weekday. It may benefit from scheduled or event-driven compute rather than twenty-four-hour capacity. A customer-facing service might require a permanent resilient base with autoscale for peaks. Storage accounts may need different retention and access-tier policies because some data is hot for a week while other objects remain auditable for years.
Cost ownership is part of architecture. If application and platform teams disagree about who funds logs, private networking or disaster-recovery capacity, expensive dependencies can remain hidden until an anomaly occurs. Establish tagging and subscription accountability, plus a process for reviewing shared-service allocation. Sustainable cost governance uses measurement and change approval, not abrupt service deletion when a budget alert arrives.
Test performance under representative load
Performance efficiency asks whether the workload meets response and throughput expectations with appropriate resources. VM CPU is one piece of evidence, but a slow request may spend most of its time waiting on a database, a remote API, DNS, or a connection pool. Scaling the web tier cannot solve every downstream bottleneck and may worsen contention on a shared dependency.
Establish user-oriented metrics: latency distributions, successful transactions, queue age, concurrency and data-processing deadlines. Test loads that resemble the application, including uneven traffic and realistic data sizes. A benchmark that exercises one endpoint with no authentication or backend calls can overstate capacity and lead to false scaling assumptions.
Scaling introduces its own reliability risks. A new instance may start before its dependencies are ready, while aggressive autoscale oscillation increases cost and complicates diagnosis. Include warm-up, health probes, state handling and minimum capacity in the test. Performance improvements should be evaluated alongside cost and reliability rather than approved solely because a graph moves down.
Make operational excellence observable in production
A service that requires its original author to diagnose every failure is not operationally mature. Teams need deployment history, appropriate logs and metrics, runbooks, recoverable configuration, safe change controls and a named responder. A dashboard alone is insufficient if nobody knows which signal warrants action or how to verify recovery.
Observability design should start with likely incidents and the questions responders will ask. Which request failed, on which version of the service, and after which change? Where did latency increase? Which dependency returned an error? Instrument those paths before collecting every possible diagnostic category at maximum retention. Redundant or noisy telemetry can become expensive without improving understanding.
Automated deployments and infrastructure as code help reduce drift, but automation without checks can rapidly replicate mistakes. Use validation, staged changes and reversible releases. Successful deployment means the service meets its health criteria after the change; an API reporting Succeeded proves only that an operation was accepted, not that the business transaction works.
Examine trade-offs as specific decisions
Architecture reviews are strongest when they compare plausible options against the workload’s requirements. For example, adding a central egress firewall can improve common policy enforcement while introducing routing complexity and additional cost. Private endpoints can reduce reliance on public network paths but require a carefully operated DNS design. Multi-region replication can improve disaster recovery while introducing more credentials, deployment configurations and recovery testing.
Write the alternatives, expected benefit, residual risk and validation evidence. Avoid generic statements such as “always use three zones” or “always use the cheapest tier.” These may be reasonable for some workloads but can misrepresent the constraints of another. A design review should also identify what has not yet been tested, so the absence of evidence is not mistaken for approval.
Revisit decisions when the business or workload changes. A service that once ran monthly may become customer-facing and need new reliability and access expectations. A platform that adds data residency constraints may require region and replication changes. Well-architected design is a maintained decision process, not a certificate permanently attached to a deployed resource.
Turn architecture review into an operational experiment
Choose one nonproduction workload and trace an ordinary user request end to end. Document its identity checks, network hops, data reads, compute dependencies and observed latency. Then model a single failure—one VM instance unavailable, one role assignment removed or one backend responding slowly—and predict the expected effect before testing it safely. Compare the result to the documented requirements.
Use the findings to prioritize improvements. A missing backup restore test may be a more serious gap than an unoptimized dashboard. A shared elevated identity may be more urgent than a small difference in storage cost. Record risk owners and due dates rather than adding a long, unordered list of recommendations that nobody is responsible for implementing.
Azure administrators can supply evidence about resource identity, connectivity, monitoring, and recovery, while application owners decide which business outcomes are acceptable. A design review should put those perspectives beside one another, with named risks, measurable targets, and a plan to revisit assumptions after the workload changes. A service is well architected when teams can explain and validate its behavior under failure, cost pressure, and ordinary maintenance.