Mental model
Classic campus networks are three-tier: access → distribution → core. That model was optimized for north-south traffic, users at the access edge talking to servers or the internet, which sat at or beyond the core. Modern applications flipped it: microservices, virtualization, container orchestration, and big-data workloads generate huge east-west traffic (server-to-server within the same DC), and three-tier’s per-tier oversubscription makes east-west slow and unpredictable.
Spine-leaf flattens the DC to two tiers. Every leaf switch (the top-of-rack, connecting servers) is directly cabled to every spine switch (the backbone). Two servers on different racks always talk over the same two-hop path: server → leaf1 → spine → leaf2 → server. Latency is deterministic. Bandwidth is easy to scale.
The layout
Spine 1 Spine 2 Spine 3 Spine 4
│ │ │ │
┌──────┼──────┬───────┼──────┬───────┼──────┬───────┼──────┐
│ │ │ │ │ │ │ │ │
Leaf-A Leaf-B Leaf-C Leaf-D Leaf-E ...
│ │ │ │ │
servers servers servers servers servers
Rules:
- No leaf-to-leaf uplinks. All east-west goes leaf → spine → leaf.
- No spine-to-spine uplinks. Spines only forward between leaves.
- Every leaf connects to every spine. With 4 spines and 24 leaves, that’s 96 spine-leaf uplinks.
- Uniform port count on spines (so all leaves have equal reach).
Why it beats three-tier for the DC
| Property | Three-tier (core/dist/access) | Spine-leaf |
|---|---|---|
| Layers between two servers | 4 hops (worst case: access → dist → core → dist → access) | 2 hops always (leaf → spine → leaf) |
| Latency | Variable: depends on where the two servers sit | Deterministic: same for any pair |
| Scaling bandwidth | Upgrade core links (expensive, disruptive) | Add another spine (parallel, non-disruptive) |
| Scaling port count | Add distribution + access | Add another leaf |
| Spanning-Tree convergence | Central bottleneck | Rarely uses STP: routing overlays instead |
| Redundancy | Distribution-tier bottleneck if one distro fails | Any single spine or leaf can fail; all others keep forwarding |
What runs on spine-leaf in practice
- Layer 3 to the edge. Every leaf terminates an L3 boundary. No VLANs stretched across leaves. This kills classic STP problems.
- ECMP (Equal-Cost Multi-Path) routing: traffic is spread across all spines. OSPF, BGP (usually eBGP-in-the-DC), or IS-IS handle the underlay.
- VXLAN overlay: L2 segments needed by tenants are extended over the L3 fabric via VXLAN tunnels between leaves, letting VMs move between racks without renumbering.
- EVPN control plane: BGP EVPN advertises MAC/IP reachability across the overlay.
A worked example: sizing a small fabric
Say you need racks for about 380 servers with 25G network cards.
- Leaves: 8 leaf switches, each with 48 x 25G server ports and 4 x 100G uplinks.
- Spines: 4 spine switches. Each leaf sends one 100G uplink to each spine.
- Servers: 8 leaves x 48 ports = 384 server ports.
Oversubscription compares what the servers on one leaf can send with what the leaf can send upward:
Server side per leaf: 48 x 25G = 1,200G
Uplink side per leaf: 4 x 100G = 400G
Oversubscription: 1,200 / 400 = 3:1
3:1 is a common design target for general compute. Storage or AI clusters aim lower (1:1 to 2:1) by adding uplinks and spines.
How it grows:
- Need more ports? Add a leaf and connect it to every spine. Nothing else changes.
- Need more east-west bandwidth? Add a spine, and give every leaf one more uplink.
- The ceiling is the spine port count: a spine with 32 x 100G ports can connect at most 32 leaves. Beyond that, large data centers add a third tier (a “super-spine”) and build several pods.
What runs on the links: every leaf-to-spine link is a routed point-to-point link (the underlay), usually with eBGP or OSPF. Because every leaf has equal-cost paths through every spine, ECMP spreads flows across all of them. If Layer 2 has to stretch between racks, a VXLAN overlay with BGP EVPN runs on top. Cisco ACI and NX-OS VXLAN EVPN fabrics both use this layout.
Leaf roles: compute, border and service leaves
All leaves have the same wiring, but they do different jobs:
| Leaf role | What plugs into it |
|---|---|
| Compute (server) leaf | Servers and hypervisors, usually one or two leaves per rack (top of rack) |
| Border leaf | The way in and out of the fabric: WAN routers, the internet edge, the campus core, other data centers |
| Service leaf | Shared appliances such as firewalls and load balancers |
Spines never connect to anything except leaves. Keeping external connections on border leaves keeps the spines simple and lets you replace them without touching the WAN.
When you’d still use three-tier
- Campus networks with heavy north-south traffic (users at access edge → internet/central servers). Three-tier’s aggregation model handles this well.
- Very small deployments: a two-tier collapsed-core is often enough.
- Environments with legacy L2 requirements where the applications assume flat VLAN behavior and can’t tolerate L3 boundaries at every rack.
The classic CCNA question
“Which topology is most common in modern data centers?” → spine-leaf.
“What kind of traffic does spine-leaf optimize for?” → east-west (server-to-server).
“How many spine switches does each leaf connect to?” → all of them (leaves fan out to every spine).
The #1 mistake
Cabling a spine-leaf like a three-tier network: adding leaf-to-leaf uplinks or spine-to-spine uplinks. This creates layer-3 forwarding loops or convergence issues and defeats the deterministic 2-hop property. The whole point is that spines never talk to each other and leaves never talk to each other directly.
Related deep-dives
- Hierarchical network design: the classic three-tier campus model
- VXLAN basics: the overlay that runs on top of the fabric
- SDN controllers: the controller model that manages large spine-leaf fabrics
Where spine-leaf shows up on US network engineer interviews in 2026
Three questions keep coming up with US employers running VXLAN/EVPN data centers (Cisco ACI, Arista EOS CloudVision, Juniper QFX, Nvidia Cumulus):
- “In a spine-leaf fabric, how are the spine switches connected to each other?”
- “You add a new rack. Do you add a leaf, a spine, or both?”
- “Explain why spine-leaf solves east-west latency and three-tier does not.”
For #1 the answer is NOT CONNECTED to each other. Spines connect only to leaves. The exam loves this trap: no spine-to-spine, no leaf-to-leaf.
For #2 adding a rack is a leaf add. You add a spine only when east-west bandwidth is saturated. Each new leaf attaches to EVERY existing spine (full mesh at the leaf-spine boundary).
For #3 three-tier has oversubscription per layer (access-to-distribution usually 4:1, distribution-to-core 2:1 to 4:1). East-west traffic between racks hits every oversubscription point. Spine-leaf has one hop up to a spine and one hop down to the destination leaf for every east-west flow. Deterministic latency, deterministic bandwidth. That is the whole value prop and it is what US DC architects pay for.
US employer context
Spine-leaf is the standard design in modern data centers: cloud providers, large enterprises and colocation tenants all use it. CCIE DC, CCNP DC, and some senior CCNP ENCOR roles ask about spine-leaf plus VXLAN/EVPN overlays. The on-prem CCNA/CCNP track touches spine-leaf conceptually but does not require you to configure one.
Related reading on this site: hierarchical network design, VXLAN basics, network topologies overview, SDN controllers (Cisco ACI, VMware NSX).
FAQ
In a spine-leaf topology, how are spine switches connected?
Not directly to one another. The whole value of spine-leaf is that every leaf is one hop from every spine, and every spine is one hop from every leaf. Spine-to-spine or leaf-to-leaf links would reintroduce the exact east-west unpredictability that spine-leaf was built to eliminate.
Is spine-leaf the same as leaf-spine?
Same architecture, different word order. Some vendors (Arista, Cumulus) say “leaf-spine”, some (Cisco) say “spine-leaf”. Resumes get both.
How many spines do I need?
Enough to carry east-west peak traffic WITH one spine failed. If a leaf has four 100G uplinks and you have four spines, losing one spine still leaves 300G of leaf-to-fabric bandwidth. Rule of thumb for small DCs: four spines. For hyperscale: eight to sixteen.
Can I run spine-leaf without VXLAN?
Yes, as a pure Layer 3 ECMP fabric. VXLAN comes in when you need Layer 2 overlays (VM mobility, multi-tenancy) on top of the Layer 3 underlay. CCNP ENCOR teaches the IP-only fabric. VXLAN overlays are CCNP DC and CCIE DC territory.
Is three-tier dead?
No. Three-tier (access, distribution, core) is alive and well in campus LANs where traffic is still north-south dominated (users to internet / users to a central app server). Spine-leaf wins in data centers where traffic is east-west dominated (server-to-server, microservice-to-microservice).
What is a border leaf?
The leaf (or pair of leaves) that connects the fabric to everything outside it: WAN routers, the internet edge, the campus network and other data centers. Routes from outside enter the fabric there, and spines stay dedicated to leaf-to-leaf traffic.
What does oversubscription mean in spine-leaf?
The ratio between the bandwidth the servers on a leaf can send and the leaf’s uplink bandwidth to the spines. 48 x 25G of server ports with 4 x 100G of uplinks is 1,200G to 400G, so 3:1.