NodeWright Updates a GPU Fleet Without Stopping the Training Run

Host maintenance on a GPU cluster has always been the job nobody wants. NVIDIA just open sourced the tool its own teams used to stop doing it by hand.

• NodeWright is a Kubernetes native package manager for the host operating system, released under Apache 2.0 after years of internal use under the name Skyhook.
• It walks each node through cordon, wait, drain, apply, interrupt and uncordon, and it respects PodDisruptionBudgets and protected workload labels.
• A DeploymentPolicy rolls changes out across labelled compartments in fixed, linear or exponential batches, stopping on a failure threshold instead of cascading.
• It manages the host layer beneath the GPU Operator, not the operator itself, so kernel parameters, security agents and CVE fixes land without replacing scarce nodes.

Read the full analysis