Currently if we see a configuration with more than one URL for signal servers, we just pick the first one.

That's because we don't currently support having more than one signal server per discovery key. (If Alice talks to signal server X and Bob talks to signal server B, they'll never be introduced even if they have the same discovery key.) That's because servers don't know anything about each other.

To support multiple signal servers, I'd like to implement a swarm along these lines:

• The server starts out with a list of known servers and tries to establish a direct websocket to each.

  • If a connection fails, it retries with exponential decay

• When any two servers connect, they sync up on an Automerge document containing a directory of known servers, along with the discovery keys that each one handles, and the number of clients currently connected for each.

• If the latest document contains servers that this server is not connected to, it tries to connect to each one.

• The client starts out with the same list of known servers, and picks one somehow (randomly, lowest ping, etc.)

  • If a connection fails, the client picks another and tries it

• When a client connects:

  • The client asks to be introduced to peers for a discovery key
  • If the server knows of another server that has already claimed that discovery key, it refers the client to that server
  • If the server hasn't seen the discovery key before, it claims the discovery key, accepts the client connection, and updates the shared directory document