Modern data centers replace the old core/aggregation/access tree with a leaf-spine fabric, a three-stage Clos network. This topology diagram shows a small fabric of four spines and six leaves, how traffic spreads over equal-cost paths, and what happens when a spine fails.
1. The Diagram
Spines. Four spine switches form the backbone. They connect only to leaves, never to each other or to servers.
Leaves (top of rack). Each rack has a leaf switch that connects its servers and has one uplink to every spine. That full mesh between the two layers is the defining rule.
Border leaves. Two leaves (green) connect the fabric to WAN routers and firewalls. External routes enter here, so north-south traffic is kept off the server leaves.
Routing. Each leaf runs eBGP to each spine with its own private AS number. Because every spine path has the same cost, the leaf uses ECMP to hash flows across all four (yellow lines from Leaf 1).
Predictable latency. Any two servers in different racks are exactly leaf-spine-leaf apart. To grow, add leaves; to add bandwidth, add spines.
2. How to Read It
In fabric diagrams the line count matters: leaves times spines links. Highlight one leaf's uplinks to show ECMP, and colour border leaves differently because their role differs even though their wiring is identical.
Also write the oversubscription ratio somewhere on the page. With 40 servers at 25G per rack (1,000G down) and 4 x 100G up, each leaf is 2.5:1 oversubscribed, which is a typical figure for general compute; AI training fabrics aim for 1:1.
3. Variant: Spine Failure
When one spine dies, BFD detects it and BGP withdraws the routes through it. Leaves rehash over the three survivors, losing a quarter of uplink capacity but no reachability. The same view helps size the fabric: with N spines one failure costs 1/N of capacity.
4. Pros, Cons, and When to Use
Pros
Shows the regular, repeating structure that makes the design scale.
Equal-cost paths and failure impact are easy to reason about.
Border leaves separate north-south from east-west traffic.
AS numbers and link speeds fit directly on the boxes.
One diagram covers dozens of racks because they all look alike.
Cons
The full mesh gets visually dense beyond 4 spines and 8 leaves.
Overlay (VXLAN tenants, VRFs) needs a separate logical view.
Cabling details (optics, ports) are abstracted away.
Can suggest symmetry that the real build lacks.
When to use
Designing or expanding a data center or GPU cluster fabric.
Explaining ECMP, failure domains and oversubscription.
When not to use
Small offices with a single switch stack.
Explaining tenant isolation (use an overlay or VRF diagram).