Live Migration

Overview

Live migration moves a running virtual machine from one physical host to another without the VM observing the interruption. In hyperscale cloud operations it is a core reliability mechanism: it is needed because hardware is imperfect, and the single biggest driver is load balancing so that large and small VMs can coexist efficiently across the fleet. As framed at the 2025 OCP Global Summit, customers "hate to be interrupted," so the migration must be seamless and unobservable to the workload.

The drivers for live migration fall into four categories: device failure that does not affect all VMs on a host, predictive failure detection using AI and machine learning, node maintenance and firmware updates, and load balancing to make room for large VMs. Mechanically, migration copies the namespace dirty map first, then drains a dirty queue; if the VM dirties blocks faster than they can be drained, the VM is throttled until the remaining state is small enough that the VM can be paused and its final state migrated. This convergence technique lets even write-heavy workloads (for example, an EC-to-DC class migration) complete migration.