That's not true though - you'll still be talking to everyone on every homeserver that yours can reach (or can be reached transitively)! When matrix.org went down for a bit a few years ago, I could talk in the same rooms with everyone on a homeserver that wasn't matrix.org flawlessly, even though the room was first created on matrix.org. I struggle to see how Matrix could be less available than IRC, where I've experienced many more netsplits.

Yes it is possible to mirror rooms, but last I checked that had to be setup manually? Or does synapse do that automatically now?

In any case that wasn't my point earlier. That was that I never had the issue on IRC that my homeserver somehow got stuck for several minutes while it did "something" in the room resolution protocol before suddenly catching up all of a sudden. And that I'll take netsplits any day over it being stuck.

I'm not sure what you mean by mirroring rooms - rooms don't have any one particular home or special server, they exist across all participating servers always and automatically.

I've never experienced what you're describing with the homeserver getting stuck while it did something, though I have had Element Android time out and seemingly not notice for a while & keep spinning, but that was the client's fault, not the server. My best guess is that whatever you experienced did get fixed!

Yes, but you are not syncing between all of them (since that would scale quite badly), but directed graph like in IRC (if I understand the specs correctly anyway). Edit: Oh.. or does it really sync between all servers (in which case I didn't quite understand the spec I guess)? Then that solves netsplits quite well at the cost of quite poor scalability.

Not sure what the issue was in the end, but I have had it with all clients I had tried (mobile, desktop, ..).

I agree that full mesh network is bad for scaling, but IRC's directed graph—manually configured even—is bad for availability. There is some talk about making the routing better, but I guess it's not in near future..

In Matrix it could be a problem, but practically it's not a problem even in the largest room #matrix:matrix.org with probably thousands of servers. But it makes joining that room quite slow and resource intensive.

Of course there is no IRC network with thousands of servers, because accepted hosts are well curated by the rules of the IRC network whereas any Matrix server can join the global federation. Granted, you don't need to join a chat network with an IRC server, but then neither does a Matrix server.

Re: edit, yep I believe it is full mesh sync for homeservers in a room, but if for some reason two can't communicate they can get the graph from other servers. It's eventually consistent if the homeservers are connected through any path, but it tries to be full-mesh. Scalability I think will be tackled to make P2P more practical, though.

Sorry you had that issue!

Incorrect, as you can configure your homeserver to be a public server or an invite-only server where you can have your chums join your homeserver - and thanks to federation even if there is some break in comms between your homeserver and another federated server - say the matrix.org server (which runs on the same matrix-synapse you'd be running) - it's possible for the federation to route around the problem - which, exactly like IRC's netsplit, will resolve eventually, but with the added advantage of synching history.

We are conflating things here, since the architecture of IRC and Matrix are not the same. Since due to the federated nature each room will be different - like it is in Jabber.

I am not talking about talking to my friends (since that's already solved with a private irc network), but for random people on rather large public rooms, where everyone will be isolated to their homeserver without manual configuration (granted.. since we have some big homeservers like matrix.org probably they'll able to continue chatting in the room).

Edit: Turns out I misremembered the federation algorithm from the spec - seems like it is indeed currently full-mesh and not a directed graph so I retract my remark.