Le mar. 16 févr. 2021 à 21:03, Job Snijders via NANOG <nanog@nanog.org> a écrit :
https://labs.ripe.net/Members/erik/ripe-ncc-and-duke-university-bgp-experime...
The experiment triggered a bug in some Cisco router models: affected Ciscos would corrupt this specific BGP announcement ** ON OUTBOUND **. Any peers of such Ciscos receiving this BGP update, would (according to then current RFCs) consider the BGP UPDATE corrupted, and would subsequently tear down the BGP sessions with the Ciscos. Because the corruption was not detected by the Ciscos themselves, whenever the sessions would come back online again they'd reannounce the corrupted update, causing a session tear down. Bounce ... Bounce ... Bounce ... at global scale in both IBGP and EBGP! :-)
In a similar fashion, a network I know had a massive outage when a failing linecard corrupted is-is lsps, triggering a flood of purges and taking down the whole backbone. This was pre-rfc6232, so you can guess that resolving the issue was a real PITA. This kind of outages fuels my netops nightmares.