Impact and scope
An operations team could misread a healthy EVPN dashboard as proof that a bridged service works. In this invented case, small keepalives succeed while full-size CE echo tests fail both ways. The affected service is IPv4 L2 bridging; IPv6 and routed gateways are outside the evidence.
Investigation record
- Signaling was stable: sessions, imported endpoint routes and installed MAC forwarding all matched the intended service.
- Access VLAN, tunnel VNI and bridge membership agreed; direct neighbor resolution and small echoes ruled out total endpoint isolation.
- Size sweeps isolated the boundary:1422-byte payload passed,1423 failed. A correlated egress MTU-drop increment tied loss to the constrained underlay rather than a speculative BGP restart.
Root cause and causal arithmetic
The modeled path IP MTU was1500, but a1472-byte ICMP payload created a1500-byte inner IP packet and a1550-byte outer IP packet. Outer DF and explicit egress drop were part of this fixture. The extra50 bytes include the14-byte inner Ethernet header plus20-byte outer IPv4,8-byte UDP and8-byte VXLAN. Outer Ethernet/FCS do not count toward IP MTU. No tags, IP options or extra tunnels were present.
Controlled repair and verification
Before any real change, confirm the complete relevant underlay path, both directions, ECMP alternatives and platform MTU conventions. The educational post-change profile uses1600-byte path IP MTU: all four probe sizes then pass both ways, oversize-drop counts stop growing, no new fragmentation appears and EVPN remains established. This is simulated evidence, not an executed configuration or a1600-byte universal prescription. Reducing the inner service MTU is another option with application and endpoint consequences.
Lessons for the incident record
Record the packet envelope, test direction and counter window with every conclusion. Separate a reachable small probe from adequate transport MTU. RFC7348 permits intermediate routers to fragment in general; the DF/drop behavior here is a declared condition, not a blanket rule.
Practice the ten-check evidence decisions in the interactive lab