There are no instances in atproto: how the protocol differs from Mastodon

Every time an article about atproto — the protocol behind Bluesky — reaches the front page of Hacker News, the comments fill with the same question: where are all the Bluesky instances? In a post published on overreacted.io on June 19, Dan Abramov answers that the question is a category error: there are no instances in atproto at all. Instances, he argues, are a Mastodon-shaped concept that does not map onto how the protocol is built.

Hosting and aggregation are two different things

The explanation starts with RSS. In the golden age of blogs, people published on their own sites, which they either self-hosted or ran on a blogging platform, while apps such as Google Reader and Feedly aggregated everyone's output. The important detail is that hosting and aggregation were separate: posts never lived inside Google Reader, and those apps were merely projections of the blogosphere. Traditional social media collapsed the two into one box with one app, which produced centralisation and runaway network effects.

How Mastodon instances work

Mastodon's answer was to make that box self-hostable, so every community could run its own little Facebook. Because each instance is a separate space with no shared source of truth, they must forward posts to each other — federation. That has consequences: your identity includes where you are from, so a Mastodon handle resembles an email address; if two instances stop federating, friends disappear from your feed; if your instance goes down, your identity ceases to exist; and the links between instances grow as O(n²).

What atproto does instead

atproto separates hosting from aggregation at the network level, much like RSS and Google Reader. There is hosting you can swap and apps that aggregate over everyone's hosting. Abramov writes that he moved his own account to a different host on the day he published, and that the process was automatic apart from three or four UX snags; a more adventurous user could host all their data themselves, for free, on Cloudflare. Apps are independent of Bluesky: he points to Tangled and Semble, and to his own small app, Sidetrail.

The post closes with two questions that actually measure decentralisation in this model: are people migrating to alternative hosting, and are people building and trying new apps? Running many copies of the Bluesky database server is possible — Blacksky does it to serve a specific moderation philosophy — but it is no more useful than running many copies of Google Reader, and network infrastructure such as relays has been cheap to operate for a year.