Microsoft Azure says failure prediction cuts hardware VM interruptions by 36%
Microsoft Azure says its Azure Failure Prediction and Detection system has reduced hardware related virtual machine interruptions by 36% across the Azure fleet. The system uses machine learning and telemetry to identify, mitigate and repair hardware failures before they disrupt workloads.
- AFPD became fully operational across Azure in 2024.
- Microsoft says hardware related VM interruptions fell by 36%.
- Monthly node repair rates increased by 80%, according to Microsoft.
- The system supports General Purpose, Specialized Compute, HPC/AI and select Storage workloads.
- Customers can receive AFPD events through Resource Health, Resource Graph, Event Grid and Scheduled Events.
How Azure Failure Prediction and Detection works
Azure Failure Prediction and Detection, or AFPD, became fully operational in 2024, according to Microsoft Azure Chief Technology Officer Mark Russinovich. It is designed to replace a reactive process in which a component fails, a workload is interrupted, and engineers investigate before a reboot or hardware replacement.
The system operates in three phases. First, machine learning models and detection rules monitor telemetry across the Azure fleet, including node behavior, component signals and indicators such as Solid State Drive health. When a node is flagged, AFPD marks it as unavailable for new workloads and, where eligible, live migrates existing customer workloads to a healthy node. Microsoft says this usually causes only a pause of a few seconds and does not require a reboot.
For cases where live migration is unavailable, Azure posts a notification with a redeploy deadline and recommended action in the virtual machine's Resource Health and Scheduled Events. The affected node is then removed from production, diagnosed, repaired with proactively staged spare parts and returned to the fleet.
Measured gains and support for SAP workloads
Microsoft says AFPD has increased the monthly node repair rate by 80% and reduced the overall number of hardware reboots by more than 36%. The system also received the Best Paper Award at the 2025 IEEE Cloud Summit for its approach to predicting, mitigating and remediating node interruptions at scale.
SAP workloads are a prominent example because they can support finance, supply chain and operations systems where brief downtime can have significant consequences. Microsoft describes SAP as a clear demonstration of AFPD's value, rather than the system's exclusive focus.
AFPD operates across Azure Compute's General Purpose and Specialized Compute offerings, HPC/AI workloads and select Storage scenarios. Microsoft says it runs automatically across the fleet and is not a separately paid tier or a capability customers must configure at the infrastructure layer.
Notification channels give customers operational options
Azure exposes AFPD findings through Resource Health, Resource Graph Explorer, Event Grid and Scheduled Events. Resource Health provides details and recommended actions for individual virtual machines, while Resource Graph Explorer allows customers to query health status across a fleet. Event Grid can send changes to monitoring and automation through Azure Functions, webhooks or REST APIs.
Scheduled Events can provide advance virtual machine notifications through IMDS, allowing configured SAP high availability clusters to move resources away from affected nodes. Microsoft says the notification content is identical across Resource Health, Resource Graph and Event Grid, so customers can select a channel based on scale and integration needs.
For users running SAP or other mission critical workloads on Azure, Microsoft recommends checking that AFPD events are being consumed through these channels and preparing automation to process Scheduled Events, including events with a NotBefore time up to seven days in the future.
