Skip to content

Azure Local Azure Monitor research

Research was revalidated on August 13, 2026 against current Microsoft documentation. Azure Monitor Health Models remains a preview service. The current resource graph uses Microsoft.Monitor/accounts/healthmodels and child resources at API 2025-05-03-preview.

Confirmed platform contract

AreaEvidenceDesign result
Health modelRoot entity, Azure-resource and generic entities, relationships, health states, signals, and state alertsUse a domain hierarchy with worst-of dependencies
IaCHealth model, authentication setting, entity, relationship, discovery-rule, and signal-definition resourcesBicep is the source of truth
IdentityManaged identity authentication settings are supportedUse system-assigned identity; grant least privilege outside the template when source scopes vary
SignalsAzure resource metrics, Log Analytics queries, Prometheus queries, Resource Health, and external health reports are product conceptsEnable only sources proven in the target environment
Azure Local metricsMore than 60 documented compute, storage, network, VM, VHD, and GPU metrics from the telemetry extensionStart with low-cardinality cluster signals; keep high-cardinality series out of default health
CollectionHealth Models evaluates existing telemetry and does not collect itPrerequisite checks must fail before deployment when telemetry is absent

Initial signal curation

SignalSourceInitial useRationale
Percentage CPUAzure Local platform metricHealth signal, parameterized thresholdsDocumented, low-cardinality when evaluated at cluster scope; duration and threshold need lab tuning
Cluster node Storage DegradedAzure Local platform metricHealth signal, parameterized thresholdsDirectly describes failed or missing drives; validate unit and aggregation behavior in the lab
Network throughput/RDMAAzure Local platform metricsWorkbook onlyWorkload-dependent; no universal health threshold
Volume/drive/VHD latencyAzure Local platform metricsWorkbook firstRequires per-hardware and per-workload baselines
VM metricsAzure Local platform metricsExcluded from infrastructure health v1Guest/workload scope and high cardinality
Health Service faultsLocal PowerShell todayFuture Log Analytics or external-health-report spikeDo not invent a cloud table or unsupported ingestion path
Registration, update, Network ATCLocal APIs todayFuture evidence spikeAdd only after a supported Azure signal is demonstrated
Telemetry freshnessAzure Monitor data sourceRequired before releaseA stale model must become Unknown or unhealthy, not remain silently green

The initial thresholds are development defaults, not universal best practices. CPU health must use sustained evaluation; memory health should prefer available memory and pressure rather than a single percentage-used value; capacity, latency, and failure thresholds must reflect redundancy and recovery time.

Remaining spikes

  1. Enumerate the live metric definitions for representative Azure Local releases and capture the exact namespace, names, dimensions, units, and aggregation behavior.
  2. Inventory Log Analytics tables and schemas created by Insights, AMA/DCR, and Telemetry and Diagnostics; save repeatable schema queries and negative results.
  3. Validate Health Models supported regions, provider registration, identity token acquisition, RBAC, cross-subscription resource access, and Service Group discovery.
  4. Inject compute, storage, network, registration, lifecycle, and telemetry-pipeline faults and record signal latency, state, alert, recovery, and Unknown behavior.
  5. Measure evaluation, alert, Log Analytics ingestion/query, workbook, and retention costs at small, medium, and large deployment sizes.
  6. Raise successor ADRs for any material API or signal-model changes.

Primary references

Released under the MIT License.