Why Your Federated Server Is Still a Single Point of Failure
"There is no central server" is true of a federated network and false of your account. The network as a whole survives any one server going down. You, specifically, do not, because your identity is a name that ends in the hostname of one machine.
That is the whole surprise. Federation removes the single point of failure for the network, and then quietly puts one back for each person. If you run that machine yourself, the person it fails for is you.
The identity is the hostname
Look at how federated systems name people:
- Mastodon:
@[email protected] - Matrix:
@andy:example.org - Email:
[email protected]
Everyone else routes to you by looking up that domain. The domain is the identity, not a label on it. Other servers do not know "Andy"; they know "whatever answers for example.org".
So when the box dies, nobody can reach you, and nobody can verify that anything new is from you either. The signing keys, the follower relationships and the room memberships were all held by that one server on your behalf.
What the rest of the network does while you are down
The good news is that other servers are polite about outages. Matrix servers queue outgoing events and retry with backoff, so a few hours of downtime usually just means a catch-up burst when you return. ActivityPub servers retry deliveries for a while and then give up on that inbox.
The bad news is that "a while" is a policy on someone else's machine, and you do not get to set it. Short outages are survivable. A dead disk and no backup is a different category entirely.
Quick detour: email already solved half of this
Hang on, why does email feel more robust? Because it has a built-in answer: MX records carry priorities, so you can name a second host that accepts mail when the first is down. SMTP also expects senders to keep retrying for days; RFC 5321 suggests giving up no sooner than four to five days.
That is a rare piece of failure design built into a federated protocol. A backup MX will hold your mail while you rebuild.
example.org. IN MX 10 mail.example.org.
example.org. IN MX 20 backup-mx.example.net.
It protects delivery only. Your mailbox, your folders and your login are still on the primary. Mastodon and Matrix have no equivalent of a backup MX at all.
What you lose when the server is gone for good
- Your account name, which cannot be recreated on another host.
- Your encryption keys, if device keys were backed up only to that server.
- Your follower list, on Mastodon, unless you exported it earlier.
- Your side of room membership and read state, on Matrix.
Mastodon has a migration feature, but it works by the old account announcing a move to the new one. That needs the old server to be reachable. If it is already dead, you can import a saved list of who you followed into a new account, but your followers are not carried across.
Matrix is similar. Rooms continue to exist on the other servers that joined them, so the conversation survives. Your identity in it does not, and a fresh account on a new server is a different user as far as every room is concerned.
What actually reduces the risk
None of this needs clever engineering. Most of it is boring, and boring is the point.
- Back up the database and the signing keys, and keep a copy off the machine. For Synapse that means the database and the signing key file; for Mastodon, Postgres and the secrets file.
- Export the portable bits regularly: Mastodon lets you export follows, followers and lists as CSV files.
- Pick a domain you control and keep it. If the domain is yours you can point it at a new host; if it belongs to the hosting provider you cannot.
- Add a secondary MX for mail, so a reboot does not bounce messages.
- Restore the backup onto a spare machine once, and time it.
The last one matters most. A backup nobody has restored is an untested assumption, and the first restore is a poor moment to learn that the key file was not in it.
Keeping the domain but changing the server
The cleverest cheap trick is to separate the name from the machine. Mastodon supports a different handle domain from the server, using a WebFinger redirect on your own domain. Matrix has a similar split through delegation with .well-known, so your user IDs say example.org while the homeserver runs on matrix.example.net.
That helps with moving hosts, because the name stays put while the service behind it changes. It does not remove the need for the data; you still need the database and keys to bring the same accounts back. It also adds another thing that can be misconfigured, and delegation mistakes are a classic reason federation silently stops working.
Where this leaves a small self-hoster
A solo server will never have the redundancy of a large instance with several staff, replicas and a paid on-call rota. That is fine, provided the plan for failure is honest: backups, a domain you hold, and a restore you have actually tried.
The network will carry on without you. The useful question is how long it takes you to come back as yourself.