Microsoft says its Azure Failure Prediction and Detection (AFPD) system uses machine learning and fleet telemetry to identify hardware problems before they interrupt customer workloads. Fully operational since 2024, the service coordinates detection, workload mitigation and hardware repair across Azure infrastructure. The company reports that AFPD has reduced hardware-related virtual machine interruptions by more than 36% and increased the monthly node repair rate by 80%.
The system is designed to move platform reliability beyond responding after a component failure has disrupted a workload. Instead, AFPD monitors node behavior and hardware signals, then acts on suspected problems before they become customer-facing incidents. Microsoft describes SAP deployments as a particularly clear example of the value, given the business impact that even brief outages can have on finance, supply chain and operational systems.
Three stages from detection to repair
AFPD combines machine-learning models and detection rules to examine telemetry across Azure’s fleet. Signals include node behavior and component indicators such as solid-state drive health. When the system flags a node, its mitigation service prevents new workloads from being allocated there. For eligible workloads, Azure can live-migrate existing customer workloads to a healthy node, generally with a pause of a few seconds and without a reboot.
Live migration is available for many, but not all, M-series virtual machine sizes, according to Microsoft. When migration is not supported, customers receive a notification with a redeployment deadline and recommended action through VM Resource Health and Scheduled Events. The affected node is then removed from production for diagnosis and repair. Microsoft says diagnostic details are attached to the node, spare parts are staged in advance, and repaired hardware can return to the fleet.
The company attributes a reduction of more than 36% in overall hardware reboots to the system. It also says the underlying work received the Best Paper Award at the 2025 IEEE Cloud Summit. These are Microsoft-reported results; the article does not provide independent measurements or a breakdown by workload type.
Monitoring and automation options for customers
AFPD’s findings reach customers through several Azure services. Resource Health provides event details and recommended actions for an individual VM, including whether live migration is available and any redeployment deadline. Azure Resource Graph can be used to query health status across a larger estate, which Microsoft says can help operators gain programmatic visibility into extensive SAP environments.
Event Grid can route AFPD events into customer monitoring and automation workflows, including through Azure Functions, webhooks or REST APIs. Scheduled Events deliver advance notifications through the Azure VM Instance Metadata Service (IMDS). For SAP high-availability clusters, customers can configure cluster automation to move resources away from an affected node in response to those events.
Microsoft says notification content is consistent across Resource Health, Resource Graph and Event Grid, allowing customers to select a channel based on operational scale and integration needs. The company recommends that customers confirm they are consuming the events relevant to their environment and ensure automation can handle Scheduled Events with a NotBefore time as far as seven days in the future. SAP customers using Pacemaker can follow Microsoft’s configuration guidance for the azure-events-az resource agent on SUSE Linux Enterprise Server or Red Hat Enterprise Linux.
Coverage extends beyond SAP
Although SAP is used to illustrate the consequences of interruptions, Microsoft says AFPD is not an SAP-specific feature. It operates across Azure Compute’s General Purpose and Specialized Compute offerings, HPC and AI workloads, and selected Storage scenarios. The company describes the system as an automatic fleet capability rather than a separately purchased tier or a feature that customers must enable at the infrastructure layer.
That distinction matters for operations teams: AFPD’s infrastructure-level mitigation does not replace customer-side monitoring, high-availability design or notification handling. Customers still need to decide how events should feed into their own workflows, particularly where applications require coordinated failover or other workload-specific action. Microsoft’s guidance points operators to Resource Health for investigation of individual VMs, Resource Graph for fleet-wide queries, and Event Grid for event-driven automation.
Microsoft has also published a technical paper on AFPD’s prediction, mitigation and remediation architecture. For enterprises and software providers running downtime-sensitive applications on Azure, the system represents a platform-managed layer of failure response, while customer notification and automation choices remain part of the reliability plan.
