Underlay Routing and Fabric Topologies
The underlay should be boring
The underlay's only job is to give every VTEP loopback reachability over as many equal-cost paths as possible. Point-to-point /31 links, a loopback per switch used as router-ID and VTEP source, and either OSPF or eBGP. eBGP is popular in fabrics because each leaf gets its own private ASN, you can filter aggressively at the spine, and BFD gives sub-second failure detection. OSPF is popular for operational familiarity. Either way, enable ECMP, raise maximum-paths so all uplinks are used, and ensure the hash includes L3 and L4 fields so overlay flows spread.
Overlay control plane choices
Two models: spines act as BGP route reflectors for the EVPN address family, or every leaf peers in eBGP with every spine for EVPN while the underlay stays IGP. Know the difference between the underlay session (IPv4/IPv6 unicast, loopbacks) and the overlay session (AFI 25 / SAFI 70, l2vpn evpn).
Traps
- Underlay reachable but overlay broken: usually a missing extended community, an RT mismatch, or a VTEP loopback that is not advertised.
- ECMP hashing that pins one flow to one link - check for polarization when uplink counts are powers of two.
- MTU: a path that only passes 1500-byte frames works for pings and fails for real traffic.
- Asymmetric routing caused by mixing routed ports and SVIs across a VSX pair.
- Forgetting BFD, then wondering why convergence is slow.
Rule of thumb
Draw two planes on scratch paper - physical/underlay with loopbacks and link subnets, and logical/overlay with VNIs and route targets - then label the given failure as 'underlay' or 'overlay'. Most scenarios collapse to that one question.