BGP, how the internet learns where to send a packet.

BGP (Border Gateway Protocol) is the routing protocol that glues the internet together: every autonomous system (AS) uses BGP to tell its neighbours which IP prefixes it can reach. Standardised in RFC 4271 (2006, originally 1989), BGP is also famously fragile; the Facebook 2021 outage, Google's 2019 leak, and Cloudflare's 2020 incident were all BGP-driven.

The public internet is divided into Autonomous Systems: about 75,000 of them. An AS is one organisation's network with its own routing policy: an ISP, a cloud provider, a university, a Tier-1 carrier. Inside an AS, anything goes: OSPF, IS-IS, static routes, mesh. Between ASes the protocol is BGP, defined in RFC 4271.

BGP's job is letting each AS announce which IP prefixes it can reach, and learning which prefixes its neighbours can reach. Each router builds a routing table by combining what its neighbours tell it, applying a long list of policy preferences, and picking a single best path per destination prefix. Then it forwards traffic that way until something changes.

The simulator below models a small internet with three Tier-1 carriers, two cloud providers and three end-user ISPs. Pick an origin AS (where the prefix lives) and a viewer AS (where the lookup is being done), and the path BGP would settle on is highlighted.

Each BGP announcement carries an AS_PATH: the sequence of ASes the announcement has crossed. When AS-X receives a prefix announcement from AS-Y, it adds Y to the front of the path before forwarding. The first ASN is the origin; each subsequent ASN is one hop closer to the receiver.

Two rules fall out. Loop detection: if your own ASN appears in a received path, drop it, you'd be creating a cycle. Path length: shorter paths beat longer ones in the BGP best-path tiebreak. The AS_PATH is both the routing data and the routing distance.

For any prefix, a router may know multiple paths. BGP runs a deterministic tiebreak: about a dozen criteria, evaluated in order until one path wins. The first few are the only ones that matter day to day:

The selection above is the protocol mechanic. Before it runs, the routes a router will consider are filtered through layers of policy. Each AS has policy on what it will announce and what it will accept: customers vs peers vs transit, prefix-list filters, RPKI validation, max-prefix limits, dampening for flapping routes.

This is where BGP gets dangerous. A single misconfigured policy can leak a route incorrectly, and because each neighbour propagates what it hears, one mistake can become a global incident in minutes. Pakistan's 2008 YouTube hijack, AS7007 in 1997 and the 2021 Facebook self-disconnection were all policy failures.

BGP does not require a prefix to come from a single AS, so multiple ASes can announce the same prefix. Routing converges to the topologically nearest one for each receiver. This is anycast, which most modern CDN and DNS networks rely on.

1.1.1.1 and 8.8.8.8 are anycast, as are most CDN edge IPs. A user in Sydney and a user in Stockholm both type the same hostname, get the same IP, and end up at different machines a few milliseconds away. Anycast acts as a coarse load balancer at the network layer. See the CDN guide.

BGP was designed in 1989 in a small, trusting world. Anyone can announce any prefix; nothing in the protocol verifies you own it. RPKI (Resource Public Key Infrastructure) is the modern fix: a chain-of-trust system where prefix holders publish signed authorisations; routers validate received announcements against the published authorisation.

Adoption stalled for years; it's now widespread among major networks. The MANRS initiative pushes the rest. Without RPKI, BGP works mainly because most operators are careful.

Cloudflare

~330 PoPs, anycast on the same IP across all of them. ~1 trillion requests/day handled by BGP-driven routing. Custom BGP daemon (BIRD-derived); peers with thousands of ASes globally.

AWS Direct Connect / Google Cloud Interconnect

Private BGP sessions between customer routers and cloud edge: lets you announce on-prem routes into your cloud VPC and vice versa. The standard pattern for hybrid-cloud connectivity.

Datacenter fabric

Modern hyperscale datacenters (Meta, Google, Microsoft) run BGP inside the datacenter as well: every Top-of-Rack switch is a tiny AS, BGP exchanges routes between them. Replaces older OSPF/IS-IS designs. Meta's published Fabric Aggregator design is the canonical reference.

Looking glasses

Public BGP visibility via projects like RIPE's RIS, RouteViews, and BGPlay. When something goes wrong on the global internet (route leak, DDoS, hijack), these are the tools researchers use to trace the path. Cloudflare publishes its own at radar.cloudflare.com.

The 2021 Facebook outage (Oct 4, 2021) is the best-known BGP cautionary tale. Facebook's edge routers withdrew the prefixes for facebook.com, instagram.com, and whatsapp.com from BGP after a misconfigured maintenance command, taking the whole property offline globally for ~6 hours. The recovery was complicated by the fact that Facebook's internal tools and even the badge readers in the office depended on the same DNS/BGP infrastructure. The published postmortem became a teaching example for change management.