Taking the benefit of the leaf and spine model using bare metal switches all racks in a datacenter will be separated by L3 meaning there will be no L2 connection amongst them. Reading further the OpenStack Platform 8 Director docs, I got the idea that interconnecting the overcloud does require L2 extension amongst all involved nodes, what breaks the benefit of using Clos on an leaf and spine model. So, can anyone points me where I can find the "best practices" for a multi rack, leaf and spine design for OpenStack deployment approach?

Thanks,
Moacir