How the data travels
The name leads to an address. What happens after that decides speed and resilience, and remarkably little of it is centrally organised.
Addresses and networks
Section titled “Addresses and networks”An IP address identifies a connection point. IPv4 writes it as four numbers
(203.0.113.10), IPv6 as groups of hexadecimal digits (2001:db8::1). IPv6 exists because IPv4,
with roughly 4.3 billion addresses, became too small.
Addresses are managed in blocks, not individually. The notation 203.0.113.0/24 describes such a
block, a subnet. The number after the slash says how many leading bits are fixed: the larger
the number, the smaller the network.
Autonomous systems
Section titled “Autonomous systems”The internet is a federation of independent networks. Each one is an autonomous system (AS) with its own number, deciding for itself which paths it offers and accepts.
Two forms of connection between them:
Peering means two networks exchange their traffic directly, usually without charging each other, because both benefit. Transit means one network buys onward delivery to the rest of the internet from another.
How does a network learn which path leads to somebody else’s block? Through the Border Gateway Protocol. Each AS tells its neighbours which address blocks it can reach and via which intermediate hops. Every neighbour then decides for itself which of the offered paths it takes and passes on.
This is a protocol built on assertion. There is no central register checking every announcement, which is why a wrong announcement can send traffic astray. There are safeguards against that, based on cryptographically signed announcements.
Unicast and anycast
Section titled “Unicast and anycast”Normally an address belongs to one place: unicast. Reaching a server in Frankfurt from Australia means paying the distance in latency, for every single packet.
With anycast, several locations announce the same address. Every neighbour in the network takes the path that is cheapest for it and thereby ends up at the nearest location. One address becomes many entrances.
Two effects that belong together:
- Proximity. The request ends where it started, not where the server happens to stand.
- Resilience. If a location disappears, its announcement stops, and the neighbours take the next best path without anyone intervening.
This is why authoritative DNS is almost always run over anycast: DNS queries are small, numerous and sit at the start of every connection. What that means for your own domains is described under Authoritative nameservers.
Why latency weighs more than it seems
Section titled “Why latency weighs more than it seems”A single request taking 300 milliseconds goes unnoticed. But loading a page is a sequence of steps waiting on each other: name resolution, connection setup, TLS negotiation, then the first content. Every step pays the distance again. Proximity therefore does not act linearly, it multiplies.