Microsoft AZ-104: Troubleshoot Azure VMs
A VM that looks unhealthy may have a compute-platform problem, a guest operating-system problem, a network path failure, or an application failure. A disciplined diagnosis identifies which boundary failed before attempting repair. Redeploying a virtual machine can erase useful evidence and leave a network, identity, storage, or guest-configuration fault untouched. A measured investigation begins with the observed failure and collects enough platform and guest signals to identify the layer at fault.
On this page
- Describe the failure without assuming its cause
- Check control-plane state and recent operations
- Use boot diagnostics to distinguish guest boot failures
- Diagnose network access from both ends
- Inspect the guest without creating more damage
- Check disks and VM agents as separate dependencies
- Walk one hypothetical incident through to verification
- Strengthen recoverability as part of closing the incident
Describe the failure without assuming its cause
“The Azure VM is down” can describe several distinct symptoms. Azure may report a failed provisioning state, the guest might be running but not accepting remote connections, or a healthy VM may host an application returning errors. Each requires different evidence and permissions. Write down the affected service, start time, last known successful operation and exactly what the user can or cannot do. A Microsoft AZ-104 VM troubleshooting task should begin with evidence from boot, network, storage, and monitoring before a disruptive redeployment is considered.
For a VM running a web service, first compare platform status, guest heartbeat where available, network connectivity and the application’s own health endpoint. A portal status of Running does not prove the application has started or can reach its database. Likewise, failed RDP access does not necessarily mean the guest cannot serve production traffic; an inbound rule, a credential or a bastion configuration may be the actual issue.
Preserve evidence before making changes. Capture relevant activity timestamps, deployment operations and error messages. Avoid simultaneous restarts, password resets and NSG edits, which can mask the original fault and complicate recovery. If the service is customer-facing, coordinate incident communications and use the documented rollback path rather than treating diagnosis as an untracked lab exercise.
Check control-plane state and recent operations
Examine the VM resource’s provisioning state, power state, instance view and recent Azure Activity Log entries. A failed extension deployment, unsupported SKU change, disk operation or quota issue may explain an apparent startup failure without requiring guest access. Distinguish resource creation errors from a guest that booted and later became unhealthy.
Azure operations can also be delayed or retried. If a VM is being redeployed, resized or recovered, repeatedly issuing new conflicting operations may extend the incident. Determine whether an operation is still active and inspect its status before initiating another. An automated pipeline may be recreating a resource after a human manually changes it, creating configuration churn mistaken for platform instability.
Check service health and platform notices when multiple resources exhibit similar symptoms. A single affected VM is usually not enough evidence to assert a regional outage. Conversely, a burst of related failures after an Azure platform event warrants coordinated investigation rather than rebuilding every individual machine. Scope and timing are vital.
Use boot diagnostics to distinguish guest boot failures
Boot diagnostics can expose screenshots and supported console output when the VM cannot complete startup or become remotely accessible. Microsoft’s boot diagnostics guidance covers supported observations for Windows and Linux guests. A boot error displayed before the network stack starts has a different remediation path from a successful login screen with blocked remote traffic.
Look for evidence of disk, operating-system or driver problems without assuming that every suspicious message is causal. Where a guest fails after updates, consider a supported repair or recovery workflow with snapshots and backups. For production systems, prefer recoverable steps and preserve the original OS disk state as appropriate before mounting or modifying it in a repair environment.
An observed screenshot of a login prompt is useful but not proof that services are healthy. After boot is restored, check agent and application logs, storage mounts, certificates and dependent endpoints. The goal is to restore the business function, not merely make a VM show Running in the portal.
Diagnose network access from both ends
Remote management and application connectivity follow ordinary network dependencies: DNS, IP addressing, routes, NSGs, any central firewall, guest firewall and listening service. A VM may have a private IP but no route from the administrator’s workstation. Another may have a route but an NSG deny on the relevant port. Diagnose the intended path rather than exposing a public management port as a shortcut.
Use effective security rules and routes for the specific NIC when available. Verify the source address that Azure sees, because corporate NAT, VPN clients and bastion paths may differ from the assumed source. If packets reach the VM but the port is not listening, focus on the guest service and firewall. If traffic never reaches it, inspect route and filtering evidence before touching the operating system.
Private DNS and service dependencies matter once the user can connect to the VM. A web application failing to reach private storage may need route and DNS analysis rather than VM replacement. Keep the difference between management-plane reachability and application-plane connectivity clear in the incident record.
Inspect the guest without creating more damage
When guest access is available, review event logs or system journals, CPU and memory pressure, disk capacity, disk I/O, service status and recent configuration changes. An operating-system disk near capacity can cause cascading failures that appear as slow CPU or unresponsive applications. A process crash may be caused by a missing configuration file after a deployment rather than by Azure infrastructure.
If direct login is unavailable, supported Azure run-command or serial-console capabilities can sometimes provide a controlled diagnostic path, depending on guest and agent state. Do not assume those options always work: some require a functional VM agent or specific configuration. Use the least intrusive supported method that can answer a concrete question, and keep a record of commands executed in production.
Troubleshooting commands can change state even when their names sound observational. A script that restarts networking or clears caches may briefly repair a symptom but disrupt other sessions. Review commands for side effects, test in a comparable environment, and obtain approval when the action is not purely read-only.
Check disks and VM agents as separate dependencies
An attached managed disk has its own provisioning and performance characteristics. A disk that is present in Azure may be unmounted or inaccessible inside the guest, while a disk that is mounted may be saturated by application I/O. Compare the Azure disk resource, guest device view and workload access before deciding whether the problem is attach state, filesystem integrity or throughput.
VM extensions depend on the guest agent and supported execution environment. A failed monitoring or script extension does not automatically mean the core VM is unusable, but it can remove required operational visibility or configuration. Examine extension statuses, guest agent logs and the most recent agent update before retrying failed installs. Blindly reinstalling extensions can overwrite useful evidence of a packaging or permission error.
Disk or extension repairs should have a recovery path. Snapshots, appropriate Azure Backup recovery and tested replacement procedures may be safer than making irreversible changes inside a failing VM. Decide whether the incident is best addressed by repair, restoration or reprovisioning based on service state and available backups.
Walk one hypothetical incident through to verification
Suppose an Azure VM hosting an internal API stops responding after a planned update. Azure reports it running, but the remote health check times out. The operator confirms that DNS returns the expected private IP, then checks routes and effective NSG rules for the health-check source. Those pass, but the guest does not listen on the expected port. The process log shows a startup failure caused by a changed configuration value.
That evidence justifies restoring the approved application configuration, not widening the subnet NSG or redeploying the VM. After the correction, verify that the API process is listening, the health check succeeds from the actual client network, and required downstream transactions complete. Compare observed service performance with a healthy baseline and retain the incident timeline for review. This is an illustrative scenario, not a claim that any real Azure VM was tested here.
Contrast it with a second scenario in which the process listens locally but the client cannot connect because a route to a centralized firewall is missing. The same user-facing symptom leads to a different fix. A good diagnosis explains that distinction and can defend why unrelated controls were left alone.
Strengthen recoverability as part of closing the incident
A repaired VM can fail again if the cause remains unmanaged. After service restoration, review monitoring coverage, backup health, update processes, capacity trends and dependencies that made diagnosis difficult. If the application cannot be safely rebuilt from configuration and data backups, document that risk and establish a plan to improve it.
Close a virtual-machine incident only after the user transaction works, the triggering condition is understood, and the change has not weakened access or recovery settings. Capture the failing symptom and the evidence that ruled out alternatives before resetting or redeploying anything. A green VM status is useful telemetry; it does not by itself prove that DNS, storage, the guest operating system, and the application are healthy.