In leaf-and-spine topologies, oversubscription typically occurs at the leaf switch.

Oversubscription is equal to the total amount of bandwidth available to all servers connected to a leaf switch divided by the aggregate amount of uplink bandwidth.

 oversubscription = total bandwidth / aggregate uplink bandwidth

For example, 19 servers with one 10 Gigabit Ethernet (10 GbE) port each create up to 190 Gbps of bandwidth. In an environment with four 40 GbE uplinks to the spine (a total of 160 Gbps) a 1.2:1 oversubscription results, as shown in Oversubscription in the Leaf Layer.

1.2 (oversubscription) = 190 (total) / 160 (total uplink)
Oversubscription in the Leaf Layer
Oversubscription at the leaf layer

You can make more or less bandwidth available to a rack by provisioning more or fewer uplinks. That means you can change the available bandwidth on a per-rack basis. 

Note

The number of uplinks from a leaf switch to each spine switch must be the same to avoid hotspots.

For example, if a leaf switch has two uplinks to spine switch A and only one uplink to spine switches B, C and D, more traffic is sent to the leaf switch by way of spine switch A, which might create a hotspot.