The Myth of the Self-Healing Internet

There's an old line from the early nineties about the network treating censorship as damage and routing around it. People still quote it, usually with a kind of smug confidence, as if it settles the matter. It doesn't. That line described design intent. Nobody wrote it while watching a transit provider on the other side of the planet have a bad night and half their prefixes go quiet.

I've spent enough years around network operator circles to have opinions about this, and the opinion is simple. The internet is not self-healing. What it has is heavy redundancy, and that's a different thing entirely. Redundancy lives in the topology; healing is a verb, and verbs need an actor.

The actor, almost always, is a person staring at a dashboard who decides something is wrong before the monitoring system agrees with them.

Multiple paths exist because people built multiple paths, deliberately, over decades. Protocols like BGP will withdraw a route and pick another if the numbers work out, and that much is genuinely automatic. Convergence, though, is not the same as a network staying usable. It can take minutes that feel much longer to whoever is paying for the outage. It can also converge on a path that's technically up and practically useless, or one fed by a filter someone three autonomous systems away configured wrong and hasn't noticed yet. The protocol has no concept of "healed" versus "barely routable." A person has to supply that concept.

What actually resolves most incidents I've seen, or heard about secondhand, is duller than the folklore suggests. Someone notices a graph bend the wrong way. They check it against other graphs that don't quite agree. They open a chat with someone at another network, someone they've talked to for years, and that history does more work than the routing protocol underneath it. A person who already trusts your judgment answers faster than a support queue ever will. Then somebody manually depreferences a session, pins a route, or calls a peer and asks them to stop announcing something, and the internet, as everyone downstream experiences it, gets quietly better. None of that shows up in the story about networks that heal themselves. When it works, nothing happens that anyone outside the room notices, which is exactly why it reads as automatic from a distance.

Why the framing matters

The self-healing metaphor is not just imprecise, it's convenient for the people who benefit most from that labor staying invisible. If the network heals itself, nobody needs to staff the 2 a.m. call. Nobody needs to budget for keeping experienced engineers around long enough to build the relationships that make a phone call at an inconvenient hour actually useful. Nobody needs to explain, in a postmortem, that the fix took eleven minutes because two people who trusted each other's judgment happened to both be awake and reachable at the same time, and no dashboard made that happen. Resilience gets billed as architecture instead of labor, and labor is the budget line that gets cut first, because it never shows up on a diagram.

I'm not arguing against the redundancy itself. Multiple providers and multiple paths genuinely cut down how often the bad nights happen, and that matters. None of it, though, replaces someone who has seen this particular flavor of weird before and knows, from a decade of smaller failures, which knob to turn.

"The internet routes around damage" is tidy. "People kept it running and one of them was awake at three in the morning" is accurate, and a lot less quotable.