Routing is the process of choosing the next hop. Inside your own network you control it; between networks it is a negotiation between around 75,000 independent organisations who mostly trust each other, which is a design decision from the 1980s that we have never managed to replace.

A router holds a table of prefixes and next hops. For each packet it finds every matching prefix and picks the longest one — the most specific. A /24 beats a /16 beats a /8 beats the default route 0.0.0.0/0. Nothing overrides this: not administrative distance, not metric, not protocol preference. Those only break ties between routes of the same prefix length.

Once you hold that firmly, several things become obvious. Why a single more-specific static route can black-hole traffic that everything else routes correctly. Why adding a /32 is the standard trick for pinning one host down a different path. And why route hijacking works.

• Static routes — you type them. Fine for small, stable, or deliberately-pinned paths. They do not react to failure, which is either the point or the problem.
• OSPF — a link-state protocol. Every router floods its local view, everyone builds an identical map of the area, and each runs Dijkstra against it. Converges fast, scales to a few hundred routers per area, cost defaults to a function of bandwidth. Area 0 is the backbone and all other areas must touch it.
• IS-IS — the same class of protocol, less common in enterprises, very common in large service providers because it scales better and is not tied to IP.
• EIGRP — Cisco's, now nominally open. Fast convergence, still widely deployed in Cisco-heavy enterprises.

The interview-useful contrast: link-state protocols share a map and each router computes its own path; distance-vector protocols share their conclusions and trust their neighbours. That difference is why link-state converges faster and why distance-vector protocols need loop prevention mechanisms like split horizon.

BGP is not really a routing protocol in the same sense. It is a policy protocol. It does not try to find the fastest path — it has no idea what fast means, no latency measurement, no bandwidth awareness. It finds a path that everyone in the chain has agreed to carry, which is a commercial question as much as a technical one.

The decision process, in order, with the ones that matter in practice marked:

The practical summary: local preference controls what leaves your network; AS path prepending is the crude tool you use to influence what enters it. If you announce the same prefix on two links and want most traffic on one, you prepend your own AS number two or three times on the other, making it look further away. It is inelegant and it is what everyone does.

• Transit — you pay someone to carry your traffic to the whole internet. Billed on the 95th percentile of measured throughput, which means you can burst for 36 hours a month for free and the shape of that billing drives real engineering decisions.
• Settlement-free peering — two networks exchange traffic directly, for free, because it saves both of them transit costs. Only covers each other's customers, never the whole internet.
• Paid peering — when the traffic ratio is lopsided enough that one side wants payment. This is what most large content-to-eyeball disputes are actually about.
• An internet exchange (IXP) — a shared switch fabric where hundreds of networks peer with each other over one port. You pay for the port, not per gigabit, which is why exchanges are the cheapest bandwidth on the internet.