It worked, then it did not, then it worked again

This is the pattern that drives people mad. Not a clean failure and not clean success. It loads, then a page hangs, then a reload is instant, then the next click sits there.

Intermittent behaviour gets read as an unstable service, which gets read as an unreliable market. The underlying mechanism is a lot more boring than that.

What is being rebuilt

An onion service maintains a small number of relays that act as its points of first contact. It publishes which ones they are. Those relays are not permanent. They come and go, and the service replaces them, republishing as it does.

A client that fetched the previous description will try a contact point that has since been dropped. That attempt fails. A fresh fetch gets the current description and works. Nothing broke and nothing was fixed.

Why it feels worse than a clean failure

A clean failure is a fact you can act on. Intermittency is a fact you cannot act on, and it invites theories. People build elaborate explanations for it. It only works in the evening. It works better with a different bridge. It works before the market gets busy.

These theories are mostly noise fitted to a pattern that is close to random. It is a natural thing for a person to do and it produces confident advice that does not survive contact with anybody else trying it.

What actually helps

1. Give a hung page a genuine chance before touching anything. Some of these resolve themselves once a fresh path is built underneath.
2. If it does not resolve, request a new circuit once rather than hammering.
3. Treat one failed page inside a working session as noise rather than as an event.
4. Only start counting failures when the whole session stops working, not when a click does.

What follows from getting it right

You stop forming an opinion about the market from network texture. That opinion tends to be sticky and it colours everything afterwards, including decisions about where to hold funds, which is a real consequence from a false premise.

You also stop taking the folklore seriously. Nearly all of the timing advice circulating about when a given market is reachable is somebody theorising about churn they could not see, and following it costs you nothing but teaches you a false model.

A note on why this site publishes no status

This case is the reason, more than any other. A service in the middle of replacing its contact points will answer one observer and not the next, at the same moment, with nothing wrong anywhere. Two honest checks a second apart can disagree.

A page that reduced that to a green tick would be describing one observer luck as a property of the service. A page that reduced it to a percentage would be worse, because the decimal point implies a measurement programme that nobody described.

So this site checks nothing, times nothing and claims nothing about reachability. It prints the addresses it was given. What happens when you try one is a fact about your attempt, not a fact this site could have known in advance.

• Back to the cases the network caused
• When a single path was the problem
• What you can actually observe