Hacker News Reader

Comment by rglullis | Hacker News Reader

> open protocol and decentralized

Open? Yes.

Decentralized? Not in practice. ATproto fundamentally requires every participating node to index/process every message being sent around the network. The whole thing about AppView can be summed up as "a global indexed view of every event from every node".

ATproto delegates all the crappy parts (identity management, moderation) to those that would like to participate in the network, but all of that is pointless if your messages are not shared with the AppView servers.

This has been Bluesky's intent from day one: by letting people host their own data at their own PDS, but by ensuring that the communication is only meaningful if passing through their indexers, they are effectively aiming to be to the "Social Web" what Google became to the old www.

View on HN

Our intent is to make an open social network that is able to scale past 100M users without breaking down, and to debind identity from providers.

It is also not necessary for AppViews to index the entire network. Partial syncs of the atmosphere are suitable for realworld usecases.

Let me present you with a use case: let's say that the New York Times (55.2 million followers on Twitter) decides to leave Musk's land and that they will fully embrace decentralized social media. Let's pretend that they learned their lesson and that they swore to never rely on someone else's platform to broadcast their own information, so they are looking to do anything they need to own the infra needed to publish all their content and reach anyone who wants to access them.

In our scenario, they don't care where their 55M followers are hosted, but they do care about giving accounts to their ~5k employees and they want to be able to follow any of the newsworthy people and institutions: politicians, entertainment celebrities, athletes, etc.

What would be the requirements and associated costs for the NYT to provide this, and how do these costs change if/when more institutions decide to follow suit?

An appview doing partial sync of the atmosphere is going to have two drivers of its cost: 1. number of users being synced, 2. number of direct users of the appview. #2 because some additional work is done for the direct users, such as computation of the following feed.

I won't have exact numbers until we actually finish that distro and deploy it, but my ballpark for 5k direct users with, say, 100 follows on average leading to 500k synced users, I'd say $100-200/mo with a max of $500. If you're not running a big instance, you're not paying big costs.

EDIT: I should clarify that the partial sync runmodel that we're developing will still operate on atproto's "shared heap" approach which means that activity outside of the selected sync-set (the 500k users) will not be visible to the NYT users. How we solve that limitation is an open question. We could introduce message-passing to notify about that activity, for instance. I think that's the most notable downside compared to a comparable AP deployment, which inherently operates on message-passing.

You absolutely buried the lede. Of course the idea is for the NYT to see all messages from its followers as well, not just their own follows. They should also be able to search and discover anyone that is on the atmosphere (after all, the first thing that journalists will want to do when investigating a person is to see their social media presence). With that in mind, what would be a real estimate for operating costs?

I'm really taking your "EDIT" as admission that ATProto will only be able to claim "actually decentralized" once it changes into something that resembles ActivityPub.

Reply on news.ycombinator.com