Operating system provisioning is a common and critical task in cluster computing environments. The required low-level operations involved in provisioning can drastically decrease the performance of a given solution, and maintaining a reasonable provisioning time on clusters of 1000+ nodes is a significant challenge. We present Kadeploy3, a tool built to efficiently and reliably deploy a large number of cluster nodes. Since it is a keystone of the Grid'5000 experimental testbed, it has been designed not only to help system administrators install and manage clusters but also to provide testbed users with a flexible way to deploy their own operating systems on nodes for their own experimentation needs, on a very frequent basis. In this paper we detail the design principles of Kadeploy3 and its main features, and evaluate its capabilities in several contexts. We also share the lessons we have learned during the design and deployment of Kadeploy3 in the hope that this will help system administrators and developers of similar solutions.