A relay went off in the middle of what you were doing

A session that was working stops working. Not slowly, and not with a message. It simply ends, and what follows is the sort of failure that arrives without an explanation.

Tor is made of machines run by volunteers. Those machines get restarted, unplugged, updated, moved and switched off, at whatever moment suited the person who owns them. Any connection passing through one at the wrong moment ends.

Why it feels so much like the service dying

Because the timing is perfect for that story. You were logged in. Things were working. Then everything stopped at once. Every element of the experience points at the destination and none of it points at a machine in a rack somewhere that you have never heard of.

There is also no notification of any kind. Nothing tells you which relay you were using, and nothing tells you it stopped. The only evidence available is the absence, and absence is ambiguous.

How to check cheaply

1. Request a new circuit and load the site again. A fresh path avoids the departed relay entirely.
2. If that works, this was the case and it is closed.
3. If it does not, try another onion service to see whether the problem is broader.
4. If nothing at all works, look at your own connection before looking anywhere else.

One minute of that sequence answers a question that people otherwise spend an evening speculating about.

What genuinely does not follow

• That the market is unstable. It was not involved in what failed.
• That your account is affected. Losing a connection is not losing a session state at the far end.
• That anything you submitted did not go through. That is a separate question with its own page.
• That the network is unreliable in general. Relay churn is a designed property, not a defect.

What follows from getting it right

Almost nothing, which is the point of this section. The reason it is written down at all is that an unexplained failure with no name attached tends to get given one, and the name it gets is usually the market.

Being able to say that a thing had no author is a skill. It stops the story from getting finished with whatever character was standing nearby, and that habit of finishing stories is what most of this site is about.

Churn is the design, not a fault in it

It is worth understanding why nobody is going to fix this. Relays are contributed. There is no contract, no uptime obligation and no central operator who could impose one. A volunteer switching off a machine is exercising exactly the freedom that makes the whole arrangement possible.

A network built the other way, from machines that could be compelled to stay up, would have somebody able to compel them. That is a worse property to have than the occasional dropped session, and the trade is deliberate.

So the right expectation is that connections are cheap and disposable rather than durable. Build one, use it, expect it to end, build another. Anybody treating a Tor session as a stable thing that ought to persist is going to be disappointed regularly.

• Back to the cases with no author
• The version you can act on
• The attribution rule