What Microsoft announced
Microsoft's Azure Observability Blog announced the general availability on November 17, 2026. The post frames the problem as monitoring application health with ease and confidence. That is the preview-era pitch, and GA carries it forward.
The feature has been around for a while. During the public preview it was completely free to use. The GA post doesn't state a price. It only points readers to the Azure Monitor pricing page and the regional availability page. Check both before you assume anything about cost or region support.
One housekeeping note. The Azure Updates page tied to this entry rendered "0 Updates found" when it was retrieved, so the listing itself couldn't be confirmed. Microsoft's dated blog post and its Learn documentation support the GA claim.
The core idea: state-based monitoring
Traditional Azure monitoring is resource-centric. A VM, a database and an endpoint each have their own thresholds and fire their own alerts. Microsoft's own criticism of that model is that one underlying fault can trigger a flood of alerts across dependent resources. A single alert also doesn't say whether customers are affected.
Health Models works differently:
- Each entity in a model gets a health state calculated from its signals. An entity can be an Azure resource or a logical component such as a user flow, a microservice or a team.
- That state rolls up through the dependencies you define, producing one state for the whole workload.
- Learn documents the states as Healthy, Degraded and Unhealthy. A Codit write-up of the preview describes the same three states.
Microsoft says this correlation reduces alert noise. That is Microsoft's description of the design. The sources don't offer independent measurements of reduced alert volume or faster incident resolution.
Building a model
Discovery
Microsoft lists four ways to populate a model:
- Application Insights topology for application-centric modeling.
- Azure Resource Graph queries for query-based resource selection.
- Service groups, including nested ones.
- Manual entities added in the designer for things like user flows or code components.
Microsoft Learn adds that discovery rules are deterministic, not AI-driven. You can combine several rules in one model. If two rules find the same resource, you get two separate entities for it, so watch for duplicates.
Relationships are created automatically only where the discovery type supports it. Application Insights topology builds them from its dependency data. For other types you add them in the designer.
Signals
Signals can come from:
- Azure resource metrics.
- Log Analytics queries.
- PromQL queries against Azure Monitor workspaces, including Managed Prometheus.
- Resource Health, as a default baseline signal.
- Recommended signals drawn from Azure service best practices and Azure Monitor Baseline Alerts.
- Existing alert rules, which you can import.
- Dynamic thresholds, which adapt to normal behavior.
A model doesn't need its own workspace. It queries your existing Log Analytics and Azure Monitor workspaces with its managed identity, so there is no data duplication or migration.
Third-party tools have no built-in connector. You either route their data into a supported workspace or push computed state through the Health Report Ingestion API. That API is still labeled preview in Learn, so it isn't part of the GA promise.
Azure Arc-enabled servers and Kubernetes can be modeled too. The model doesn't connect to them directly. It reads whatever telemetry they send to supported Azure Monitor sources.
Views, history and alerts
- The Graph view shows current health and dependencies.
- The Timeline view shows when an issue began, which regions and components were affected, and how long it lasted.
- Health history is kept for 30 days, and you can read it through the Health Models REST API.
- Health alerts are entity-state alerts that fire when an entity goes to Degraded or Unhealthy. They use the standard Azure Monitor pipeline and action groups. ITSM connectors, webhooks and Logic Apps work as they do for any other Azure Monitor alert.
- Alert processing rules can suppress notifications during planned maintenance.
Infrastructure as code
Health models are Azure resources under the Microsoft.CloudHealth/healthmodels provider. You can define them with Azure CLI, Azure PowerShell, Bicep, ARM templates or Terraform, and keep the definitions in source control. Learn also notes that GitHub Copilot can draft template definitions from the published schema. It says to validate the output before deploying, because only you know which services matter and what "healthy" means for them.
Getting started in the portal
The documented path is Azure Monitor → Health models → Create.
Prerequisites
- An active Azure subscription and at least one resource to monitor.
- Telemetry already available for the signals you plan to use.
- At least Contributor on the resource group, directly or inherited from the subscription.
- Owner or Role Based Access Control Administrator, to assign roles to the model's managed identity.
Steps
- In the Azure portal, search for and select Health models, then select Create.
- On the Basics tab, choose a subscription and resource group, name the model, and pick a supported region.
- On the Identity tab, keep the system-assigned managed identity enabled. You can optionally add a user-assigned identity.
- Optionally add tags, then choose Review + create, and Create once validation passes.
- Microsoft says deployment typically takes under a minute.
- Open the resource and confirm that Provisioning state shows Succeeded.
- Under Health, open Designer and confirm the root entity exists.
Roles are not automatic. This is the most likely place to trip up. The model's managed identity needs:
- Reader on the Azure resources it represents.
- Monitoring Reader on any Log Analytics or Azure Monitor workspace it queries.
Microsoft states plainly that Azure doesn't assign these roles for you. If signals show no data, check the role assignments first. Models can span resource groups and subscriptions, as long as everything is in the same Microsoft Entra tenant.
To manage a model you need Contributor on it. To view one you need Reader.
Practical guidance and limitations
- Plan for resilience. Microsoft recommends deploying the model in a different region and subscription from the resources it monitors. That keeps it available if an outage hits the workload. This is a recommendation, not a requirement.
- There is no prebuilt template. Discovery plus recommended signals gets you a first draft. You still decide which components matter and how failures affect their parents.
- Rollup can be tuned. Aggregation settings on a parent let you require a minimum number of healthy children, or tolerate a percentage of unhealthy ones. Without that tuning, one leaf failure can turn the whole workload red. Microsoft warns that poor entity choices, noisy signals, bad thresholds or overly aggressive rollup can undermine a model, and recommends starting small and iterating.
- Refresh intervals matter. Each signal has a refresh interval, typically 1 minute. Match it to how often the underlying data updates. A metric that reports hourly gains nothing from one-minute polling.
- Known gaps:
- There is no direct Grafana visualization integration for health models. You can build dashboards from the underlying Azure Monitor data, or read state through the REST API.
- There is no dedicated maintenance mode. Use alert processing rules instead.
- You can't apply Boolean logic between signals on one entity. Use child aggregation settings.
- Resource Health isn't available for every resource type.
- If a workspace restricts queries with Private Link Scope or a network security perimeter, test that the model's queries succeed.
Where it fits in Microsoft's broader observability push
A second Microsoft post from Ignite 2026 lists Health Models among several Azure Monitor announcements. These include cross-agent debugging and Azure Monitor traces with Trace Analytics, both in public preview. The post says tracing explains what happened and where, while Health Models helps operators judge customer impact and prioritize across supporting services.
Microsoft also says the Grafana MCP endpoint lets agents query Health Models alongside other data. Calls run with the signed-in Entra ID user's permissions. Health Models itself is the health-modeling capability, not an AI discovery feature.
Analysis
The sources point to two things that matter:
- The tooling is not the hard part. The designer, discovery and IaC support remove a lot of the work. The difficult step is deciding what "healthy" means for your service and which dependencies matter. Microsoft's own FAQ says as much. Teams that skip that thinking will get a tidy graph that is wrong.
- Treat the model as a living artifact. Microsoft suggests starting small and iterating. A DevOps Masterminds write-up from the preview period makes a similar point, comparing models to architecture diagrams and runbooks that need regular review.
Check the pricing page before rolling this out widely. Free preview usage doesn't tell you what GA will cost, and the GA announcement doesn't say.
References
- (Launched) Generally Available: Azure Monitor Health Models Azure Updates · 2026-11-17T20:00:32Z
- Azure Monitor Health Models is Generally Available | Microsoft Community Hub techcommunity.microsoft.com
- Azure Monitor health model portal quickstart - Azure Monitor | Microsoft Learn learn.microsoft.com