Azure Failure Prediction and Detection System Reduces VM Interruptions by 36%

Azure Launches New Failure Prediction System

Microsoft Azure has introduced its Azure Failure Prediction and Detection (AFPD) system, which became fully operational in 2024. This end-to-end system is designed to identify and address hardware failures proactively, before they cause interruptions to customer workloads. AFPD utilizes machine learning models and continuous telemetry monitoring across the Azure fleet to predict potential issues.

Significant Reduction in VM Interruptions

Since its implementation, AFPD has led to a 36% improvement in preventing hardware-related virtual machine interruptions across the entire Azure infrastructure. This reduction is attributed to the system's ability to detect degrading components and patterns indicative of future failures, such as Solid State Drive (SSD) health.

How AFPD Operates

AFPD replaces the traditional reactive approach to hardware failures with a three-phase process: prediction and detection, mitigation, and remediation. Machine learning models and detection rules continuously monitor telemetry, node behavior, and component signals. Once a potential failure is flagged, the mitigation service marks the node as non-allocatable for new workloads and, where possible, live-migrates existing customer workloads to a healthy node with minimal disruption, typically a pause of a few seconds.

Broad Workload Coverage

While SAP workloads are highlighted as an example of AFPD's value for mission-critical services, the system is not exclusive to them. AFPD operates across a range of Azure workload types, including General Purpose, Specialized Compute, HPC/AI, and select Storage workloads. This broad application demonstrates Azure's approach to enhancing reliability for various third-party services hosted on its platform.