Internet exchanges have presented as shared layer 2 networks for thirty years. I will show the mechanisms that silently blackhole exchange traffic or forward it to the wrong port: missing or incorrect neighbour entries, and unknown unicast flooding. In each case BGP stays established to the route servers and nothing reports a fault.
A large exchange stated in writing that it gives no guarantee traffic sent to a peering LAN goes to, and only to, the port of the network it was meant for, and that it knows of no exchange that does. I do not know how many members expect that answer.
The rest of the industry has spent years dismantling large layer 2 domains in favour of routed ones. Perhaps it is time for internet exchanges to forward at layer 3, rather than running a layer 2 forwarding domain with a route server making routing decisions.
In many cases this needs no new hardware. The switches already used to build exchange fabrics forward at layer 2 and at layer 3 at line rate. Route scale is where careful consideration is needed.
The shape of a layer 3 internet exchange switch is already commonplace: each PNI switch a CDN or ISP operates is a routed exchange with several settlement-free peers and one customer on it.
Running at layer 3 gives participants complete isolation between ports, the same model that already applies to every other upstream connection that networks buy. RTBH works at the exchange rather than depending on a peer to honour it. Debugging becomes possible from the participant side, because traceroute shows the exchange hop.
Licensed to the public under http://creativecommons.org/licenses/by/4.0