[AKSF_Ackermann]

I wonder why were relays mentioned only mentioned in passing and there was no elaboration on how they interact with the rest of the network. Maybe because doing that would show that there are in fact "instances" in atproto, but who knows?

[danabramov]

Author here!

I mostly skipped over it because a Relay is an optimization and not essential to the shape of the network. It's not a fundamental element in the same way that PDS (hosting) and AppViews (app servers) are. It's more like a "next reasonable thing" an engineer would bolt on to make it easy to create apps.

An app can work without a Relay (like https://reddwarf.app/ does). There are caches like Constellation (https://constellation.microcosm.blue/) that you can just query directly.

A Relay is not an "instance" in any meaningful sense because it is a dumb retransmitter. It is cheap to run one, and it is easy to pool them between multiple apps. (Fun fact for nerds: the Relay's API for subscriptions is literally the same as a single server's. So a Relay is kind of a facade for "a bunch of servers" that lets you listen to their events combined.)

Early on (more than a year ago), running a Relay used to be more expensive because any Relay was expected to store the entire network archive. This is no longer a part of the contract, but a lot of discussions still reference or assume that. The current cost of running your own Relay (if you don't want to pool with anyone) is about $30/month. There are community-run Relays like https://firehose.network/ that you can use too.

[chokolad]

> Maybe because doing that would show that there are in fact "instances" in atproto, but who knows?

I wonder why you are vagueposting here instead of stating your position firmly. Maybe because you are afraid to be shown wrong, but who knows ?

[uberex]

I just searched up Relay and I guess Relay is like Feeder in the analogy? I don't need it? And if I want it I choose one (many?) or run my own.

[hooverd]

You need a relay unless you want to have all the same NxM scaling problems as Mastodon.

[neko-moe]

Not really, considering there aren't that many app servers. Each PDS sends out its events, those events are picked up by whoever wants them, mostly relays but some apps might directly subscribe. The more apps in play, the more connections, but it is not the case that for every new PDS, every other PDS now has to open an additional connection. It scales with the amounts of apps, not PDSs.

[hooverd]

Mostly because having a big centralized firehose relay is less decentralized than people want to admit.

[danabramov]

That's not true.

It's not centralized in any way that matters. The Relay Bluesky uses is open source, you can run your own for $30/month if you really insist on doing that, it's trivial to pool between multiple apps if you want to lower costs, and there are already a few independent ones you can just use directly now, for example:

- https://pdsls.dev/firehose?instance=wss%3A%2F%2Fatproto.afri...

- https://pdsls.dev/firehose?instance=wss%3A%2F%2Feurope.fireh...

[chokolad]

Why is that a problem if you can host relay for $50 per month as TFA mentions?