How can Lemmy scale?

HelloLemmySup@sh.itjust.works to Lemmy@lemmy.ml – 3 points –

I like the idea with Lemmy/kbin and the fediverse but theres something I dont understand perhaps.

If in the future Lemmy is very popular and someone wants to add their own server and federate with everyone then from that moment that new instance will get all new comments, posts, etc. from all other instances its federated with and must save them in its db. This means if Lemmy gets popular forget about little guys helping out spread the “load” because every intance still must take and save all new data. Thats a lot of processing power and storage. How can this work? I see in the future only a few instances will survive.

If somehow each instance was a node and only took care of its posts and comments and forward them to others upon request I can understand scaling but this is not how it works AFAIK. Another way would be with consensus algorithms where a node saves more thsn its own data but still not all.

6

Individual servers do not handle all data. They only handle data required by its users.

You've misunderstood. Every instance does not contain all content from every other instance. Only that which at least one user has specifically requested by entering the id of a community in the !name@instan.ce format in search.

This means that the star trek instance, will only ever need to mostly host start trek content. It wont get flooded with everything else on the entire network, as it grows. Some portion of it, yes, as users on the star trek instance will inevitably sub to at least some stuff outside it, too.

Additionally, pictures and media are cached, but not permanently federated. When you upload a picture, you may have noticed the link becoming one that points to the instance you're posting from. This doesn't change even when that post gets federated to other instances, they are still fetching that image from the instance it was posted from (unless its a recent post, in which case the image may well be cached, as well).

This means that whats gets federated, is mostly just a bunch of text data, and even then, just a subset that is needed. A much lighter load.

At the smallest scale, you could have a node with just one user, perhaps that user creates a community or two. But this means that that instance will ONLY EVER store the subs of that one user, and the content of the communities they created. Not even close to the total content of the entire fediverse.

Ok thats a bit better. I didnt know about that detail.

Still that only moves the problem to the future. As I understand you should pick a community at random to sign up and then from that community access the rest. Then its a matter of time that enough users from StarTrek that have signed up there subscribe to enough big communities for the problem to appear, no?

Yea, I think you’re right. Once any instance has enough users with enough interests and subscriptions to enough communities, you get a scenario where a good portion of the whole network is duplicated on every or many nodes of the whole network. This is how the fediverse works, and I’ve yet seen anyone seriously address what this looks like at large scales and long timelines.

Storage space isn’t too expensive I guess, so maybe it’s something we can just solve when we come to it.

But, the problem may be worse with threadiverse platforms (lemmy/kbin and any other grouped or threaded platform) for exactly the reason you highlight … the whole community and all of its discussions get duplicated. For microblogging platforms, things are more granular as it’s only single posts by people who are followed that duplicated.

It may not be fatal and may be something we can solve when we get, which makes sense as getting up to a significant scale of users is tough in its own right … but it’d sure be nice to see someone think through the numbers.

This is literally how the entire internet works. You are describing CDNs.

Additionally, from the perspective of the protocol (ActivityPub), there is no such difference which you are describing.

Communities are "users" which can be "followed" (subsribed to) by other "real" users. Essentially they are bot users that other users can post content through, to its followers. There is nothing different in how the threadiverse functions compared to the fediverse at large. Only its format.

Except we’re not CDNs. Instances aren’t run by companies, most of the time, but by volunteers. Not that I have anything against companies running instances, but it’s not what the fediverse is about.

So the question of resources is a sensible one, whether or not the current protocol and architecture has worked in the past.

And the threadiverse’s difference in format is precisely the difference in highlighting. Sync for microblogs is over the posts of individual users that are followed. Sync for the threadiverse is over a community which comprises many users’ posts. Communities with threads versus single user microblogs … this is the format difference. And it’s the difference in what gets synced. Right, please correct me if I’m wrong.

And so, if I’m right, the question of how much gets duplicated also differs.

Whether the threadiverse has more duplication depends on the details, of course. My reasoning was that it would be easier for more duplication to occur on the threadiverse, as whole collections of conversations of many users will duplicated simply from one users single subscription. This is compared to the microblog platforms where users often only follow hundreds of people (my impression only).

Of course, it may be that any users output is distributed over many communities, so that communities turn out not to be larger overall (maybe this was your point). And also, as you say, cacheing and duplicating is how the internet works, so we should have ways to handle it.

All in all though, it would be nice to have some basic numerical analysis done, especially if we want people to start instances without worrying about getting burnt by ballooning costs.