What is the long-term storage plan for Lemmy instances?
Over time, Lemmy instances are going to keep aquiring more, and more data. Even if, in the best case, they are not caching content and they are just storing the data posted to communities local to the server, there will still be a virtually limitless growth in server storage requirements. Eventually, it may get to a point where it is no longer economically feesible to host all of the infrastructure to keep expanding the server's storage. What happens at this point? Will servers begin to periodically purge old content? I have concerns that there will be a permanent horizon (as Lemmy becomes more popular, the rate of growth in storage requirements will also increase, thereby reducing the distance to this horizon) over which old -- and still very useful -- data will cease to exist. Is there any plan to archive this old data?
Pictrs 0.4 recently added support for object storage. This is fantastic, because object storage is dirt cheap compared to traditional block storage (like a VM filesystem). This helps a lot for image storage, which is a large part of the problem, but it's not the whole problem.
I know Lemmy uses Postgres for everything else, but they should really invest time into moving towards something more sustainable for long term/permanent hosting. Paid Postgres services are obscenely upcharged and prohibitively expensive, so that's not an option.
I'm armchair architecting here so I'm not sure what that would look like for Lemmy (Cloudflare KV? Redis?)
Still, even my own private instance has been growing at a rate of about 700MB per day, and I don't even subscribe to that many things. I can't imagine what the major instances are dealing with. This isn't sustainable unless we want to start purging old data, which will kill Lemmy long term.
EDIT: Turns out ~90% of my Lemmy data is just for debugging and not needed:
I'm not really sure that a K/V service is a more scalable option than Postgres for storing text posts and the like. If you're not performing complex queries or requiring microsecond latencies then Postgres doesn't require that much compute or memory.
People can get unnecessary scared of relational databases if they've had bad experiences with databases that are used poorly, but attempting to force relational data into a K/V can lead to the application layer essentially just doing a less efficient job of the same types of queries that the database would normally handle. Maybe there'll be some future need to offload post and comment bodies into object storage or something but that seems incredibly premature.
Object storage for pictrs is definitely a fantastic addition, though.
The largest table holds data that is only needed by Lemmy briefly. There is a scheduled job to clear it... Every 6 months. There are active discussions on how best to handle this.
On my instance I've set a cronjob to delete everything but the most recent 100k rows of that table every hour.
I saw that issue, and then I saw people having problems after clearing it, so I'm just going to wait until they figure that out in a stable version. Looking forward to it though!
There is a good writeup on how to do the migration here. I went through it myself since I host my tiny Lemmy instance on an AWS EC2 instance. It went pretty smoothly bu obviously larger instances will have to take a longer downtime to perform the migration.
By the way, how are your costs on EC2? My understanding is that hosting on EC2 would be cost prohibitive from data transfer costs alone, not to mention their monthly rates for instances are pretty much always below the cost of a VPS.
Now if only someone could do this for the Postgres data. I wonder if S3FS would be able to handle the load of a running database, that would be a nice way to save costs.
It would be greater if it can also leverage IPFS. So we can have unique identifiers per media object and hence deduplication in a P2P network which in my opinion is more federvise affinitive. I have been thinking of making such an alternative media backend for a while.
AWS Postgres instances aren't that expensive, and they handle upgrades and backups for you.
That said, I'm interested in distributed storage, and maybe this fall/winter when I get some time off I'll try making a lemmy fork that's based on a distributed hash table. There are going to be a ton of issues (i.e. data will be immutable), but I have a few ideas on how to mitigate most of the issues I know about.
Just think of all those old, helpful forum posts from years past with tinypic and Photobucket links that are dead. I agree memes can probably die out after time but anything informative would be bad to lose imo
For large instances pictures is probably the bigger consumer of space, but for small instances the database size is the bigger issue because of federation. Also, mass storage for media is cheap, fast storage for databases is not. With my host I can get 1TB of object storage for $5 a month. Attached NVMe storage is $1 per month per 10 GB.
For my small instance the database is almost 4x as large as pictrs, and growing fast.
Premature optimization is not good. Content here is not very storage intensive, so I would not yet make it to issue. Postgre can handle billions of rows when indexed right.
One way to approach the geometric storage growth would be to not cache everything everywhere all at once. With 1000+ instances, storing an object in a few instances would be ok if others can pull it in on demand. Can use some typical caching methodology like use frequency, aging etc.
This is a great idea. Instances will need eventually to agree to common storage areas, even if they dont all allow the same content on their instance. That savings would be huge in the long run.
The long term solution is something like IPFS object storage that's read only for everyone but the author instance. One copy of the data but all instances can read it and it's stored forever in a redundant medium with bitrot protection.
Personally I think we should add a differentiation between the storage policies of content which is owned by your own instance and content that federates from other instances.
The former should be kept for a long time (forever?), while the latter can be cleared more regularly.