Data Center Leaf-Spine Architectures
What to know
Leaf-spine is a two-tier Clos fabric: every leaf connects to every spine, and spines connect only to leaves. This gives predictable, equal-cost paths between any two leaves (ECMP) and scales horizontally by adding spines (bandwidth) or leaves (ports). Candidates must know: why east-west traffic dominates modern data centers, how ECMP load-balances flows across spines, and the role of Layer 3 to the leaf (routed fabric) versus Layer 2 stretching.
Understand oversubscription ratios (downlink bandwidth vs uplink bandwidth) and how to compute them. Know the difference between a leaf-spine fabric and legacy three-tier (core/aggregation/access) designs, and why the two-tier Clos avoids spanning tree by routing at Layer 3. Understand MLAG for dual-homing servers and its limits versus EVPN multihoming.
Traps
- Assuming spines connect to each other; in a pure leaf-spine they do not (that would be super-spine territory).
- Confusing oversubscription direction: it is (sum of host-facing bandwidth) / (sum of uplink bandwidth).
- Thinking ECMP load-balances a single flow across many paths; by default a flow hashes to one path unless flow-label or per-packet hashing is used.
- Forgetting that MLAG is a Layer 2 dual-homing technique and does not scale like EVPN multihoming.
Worked example / rule of thumb
A leaf has 48 x 25G downlinks (1200G) and 4 x 100G uplinks (400G). Oversubscription = 1200/400 = 3:1. Rule of thumb: keep oversubscription low (1:1 to 3:1) for latency-sensitive east-west workloads; higher ratios are acceptable for north-south-heavy traffic.