01 / THE SERVER-FACING EDGE
Meet the leaf layer
A leaf switch sits at the server-facing edge of the fabric. Servers and other endpoints normally connect to leaf switches; the leaf then carries their traffic into the wider network. Depending on the design, a leaf may serve one rack or several.
In many common deployments, a Top-of-Rack (ToR) switch functions as a leaf. The terms describe different ideas: ToR describes where switching is deployed, whileleaf describes its role in the topology. They are related, but not always identical.
SERVERSLEAF
Revisit Top-of-Rack networking for the rack-level view, or see how a server NIC, Ethernet, and fiber provide the physical connection.
02 / THE INTERCONNECTION LAYER
Meet the spine layer
Spine switches interconnect the leaf switches. In a classic spine-leaf fabric, each leaf connects to each spine. Servers normally connect through the leaf layer rather than directly to spine switches.
This is a common model, not a claim that every data center uses the same topology.
THE LOGICAL TOPOLOGY
Every leaf reaches every spine
This simplified fabric has three leaves and two spines. Read each row: every leaf has a connection to both spine switches. The table labels the links directly, so the relationship stays clear without relying on line color.
| Leaf layer | Spine 1 | Spine 2 |
|---|---|---|
| Leaf 1 · Servers | Connected | Connected |
| Leaf 2 · Servers | Connected | Connected |
| Leaf 3 · Servers | Connected | Connected |
On a small screen: Leaf 1 → Spine 1 and Spine 2; Leaf 2 → Spine 1 and Spine 2; Leaf 3 → Spine 1 and Spine 2.
FOLLOW EAST-WEST TRAFFIC
How does traffic move between racks?
Suppose Server A and Server B attach to different leaves. Traffic can travel from Server A to its leaf, cross a spine, reach Server B’s leaf, and then arrive at Server B. This is a conceptual east-west path between systems inside the data center.
- SERVER ASource endpoint
- LEAF ASource rack edge
- SPINEInterconnects the leaves
- LEAF BDestination rack edge
- SERVER BDestination endpoint
Connecting every rack directly to every other rack becomes difficult to scale and manage as an environment grows. A structured fabric separates endpoint/rack connectivity at the leaf layer from inter-leaf connectivity at the spine layer.
FROM DIAGRAM TO PHYSICAL SPACE
A fabric at data-center scale
Real facilities may contain many server racks, many leaf switches, multiple spine switches, structured fiber routes, and extensive upstream infrastructure. The logical diagram explains relationships; a photograph cannot prove which topology a facility uses.
As the fabric grows, multiple spines can give a leaf more than one possible route to another leaf.
MORE THAN ONE POSSIBLE PATH
Why have multiple spines?
With multiple spines, traffic from Leaf A to Leaf B may be able to use Spine 1 or Spine 2. Modern fabrics can use ECMP, short for Equal-Cost Multi-Path, to use multiple equal-cost routes instead of depending on one path. This is the beginner idea; ECMP is not a guarantee that every path is always available or equally useful.
More possible paths can support scale and resilience, but do not by themselves guarantee service availability. Topology, forwarding behavior, configuration, remaining capacity, endpoint connectivity, and application design all matter.
TRACE THE DEPENDENCIES
What happens when a switch fails?
Spine 1 fails
Spine 1 — FAILED
Spine 2 — HEALTHY
If the topology, forwarding behavior, and remaining capacity support it, traffic may still have a path through Spine 2. Some traffic or performance may still be affected.
A leaf fails
Leaf 1 — FAILED
Other leaves and spines — HEALTHY
Endpoints that depend only on Leaf 1 may lose network connectivity even while other racks and spine switches remain healthy.
These are possible scopes of impact, not universal outcomes. The actual result depends on topology and configuration.
MULTIPLE SWITCHES ≠ AUTOMATIC REDUNDANCY
Ask what path remains
A topology that looks redundant does not guarantee resilient service. Check the dependencies:
- Does each server have redundant network attachment?
- Are there multiple leaf paths, and are they independent?
- Are multiple spines available, with enough surviving capacity?
- Does forwarding recover appropriately after a failure?
- Are there hidden shared dependencies?
Complete Network 101 with the final lesson:Network Redundancy in a Data Center.
THREE THINGS TO REMEMBER
Carry this mental model forward
- Leaf switches connect servers and racks into the fabric.
- In a classic spine-leaf design, every leaf connects to every spine.
- Multiple paths can help with scale and resilience, but topology alone does not guarantee service availability.
PUT THE DEPENDENCY MODEL TO WORK
Build a data center and trace its systems
The DataCenterTwin workspace currently models facility power and cooling. It does not simulate network topology, ECMP, or leaf/spine failures.
PASS IT ON