[devinivy]

On the upside, atproto "instances" (we often call them "personal data servers") have much less responsibility than the typical Mastodon instance. So folks should find them quite cheap and easy to run. One reason is that they don't need to directly support application features such as timeline construction, search, image optimization, etc: they just host your social data.

We have 10 servers federating in production, each housing around 270,000 users, and it all runs on disk and sqlite, quite affordable! We migrated 2 million users off a single large host onto these 10 smaller ones transparently and without need for much fanfare. We're dedicated to the tech and seeing it work in production has been super heartening.

[pfraze]

the entire architecture is like if you started with the typical architecture for social network, but then got rid of the organizational perimeters. the atproto instances (PDSes) handle primary data storage, key management, and request routing. they're otherwise quite dumb, which as divy is saying makes them cheap to run and also general purpose for other kinds of applications.

the app-specific business logic lives in the aggregator services that pull from those instances