Microsoft says its Azure Failure Prediction and Detection system, or AFPD, has reduced hardware-related virtual machine interruptions across the Azure fleet by more than 36%. The system combines machine-learning models, detection rules and fleet telemetry to identify potential hardware failures, then coordinates workload mitigation and repair before a failing node causes an interruption where possible.
AFPD became fully operational in 2024, according to Microsoft Azure Chief Technology Officer Mark Russinovich. The company describes it as a platform-wide capability rather than a feature customers must purchase or enable at the infrastructure layer. Its coverage includes General Purpose and Specialized Compute, high-performance computing and AI workloads, and selected storage scenarios.
From detection to repair
The system follows three stages. In the first, models and detection rules monitor node behavior, component signals and hardware indicators such as solid-state drive health. Those signals are used to identify patterns associated with active or likely future failures.
When a node is flagged, AFPD’s mitigation service prevents it from receiving new workloads. For eligible virtual machines, Azure can live-migrate existing customer workloads to a healthy node without a reboot. Microsoft says the migration typically involves a pause of a few seconds. Live migration is supported for many, but not all, M-series virtual machine sizes.
If live migration is unavailable, Azure posts a notification with a redeployment deadline and recommended action through the virtual machine’s Resource Health information and Scheduled Events. The affected node is then removed from production. Technicians receive diagnostic details intended to avoid repeating troubleshooting steps, and Microsoft says proactively staged spare parts help with repair before the node returns to the fleet.
Microsoft reports that AFPD has increased the monthly node repair rate by 80% and reduced overall hardware reboots by more than 36%. The company also says the system’s underlying research received the Best Paper Award at the 2025 IEEE Cloud Summit. These are Microsoft-reported results; the company did not provide additional methodology or a measurement period in the material describing the figures.
Visibility and automation for customers
AFPD’s signals are made available through several Azure services, giving customers options for reviewing individual virtual machines or integrating events into broader operational workflows. Resource Health provides details and recommended actions for a particular VM, including whether live migration is available and any redeployment deadline.
Resource Graph Explorer enables queries across a fleet, which Microsoft identifies as useful for large SAP environments. Event Grid can deliver health changes to customer automation, including Azure Functions, webhooks and REST APIs. Scheduled Events supplies advance notifications through the Azure Instance Metadata Service (IMDS), allowing configured high-availability clusters to move resources away from affected nodes.
Microsoft says the information is consistent across Resource Health, Resource Graph and Event Grid. Customers can therefore choose a channel based on how they monitor or automate their environment, rather than expecting different event details from each service. Event Grid notifications can also be routed onward to services such as Logic Apps, Event Hubs or Azure Monitor alerts.
For SAP high-availability deployments, Microsoft recommends preparing automation to handle Scheduled Events, including events with a NotBefore time of up to seven days in the future. It also points SAP customers to guidance for configuring the Pacemaker azure-events-az resource agent on SUSE Linux Enterprise Server and Red Hat Enterprise Linux.
SAP illustrates the reliability case
SAP workloads provide a prominent example because they often support finance, supply chain and operational processes where interruptions can have significant consequences. Microsoft says AFPD is not SAP-specific: SAP is an illustration of the system’s potential value for workloads with strict availability needs, while the same capability operates across other supported Azure workload categories.
The approach shifts some failure handling from customer-designed recovery procedures to platform-level detection and mitigation. It does not replace customer architecture, notification configuration or automation. Enterprises running critical services still need to decide how to consume health events and how their applications should respond, particularly when live migration is not available.
Microsoft advises customers to review their notification setup: use Resource Health to investigate a VM, Resource Graph to find affected machines across subscriptions, and Event Grid when event-driven automation is required. For SAP clusters, Scheduled Events integration can help configured systems respond to advance notice. The operational benefit therefore depends not only on Azure’s ability to detect and handle hardware issues, but also on whether customers have prepared their monitoring and response paths.