It doesn't. The usual approach is to create new nodes with the updated OS, migrate all workloads over and then throw away the old ones