Just thought I’d share this since it’s working for me at my home instance of federate.cc, even though it’s not documented in the Lemmy hosting guide.
The image server used by Lemmy, pict-rs, recently added support for object storage like Amazon S3, instead of serving images directly off the disk. This is potentially interesting to you because object storage is orders of magnitude cheaper than disk storage with a VM.
By way of example, I’m hosting my setup on Vultr, but this applies to say Digital Ocean or AWS as well. Going from a 50GB to a 100GB VM instance on Vultr will take you from $12 to $24/month. Up to 180GB, $48/month. Of course these include CPU and RAM step-ups too, but I’m focusing only on disk space for now.
Vultr’s object storage by comparison is $5/month for 1TB of storage and includes a separate 1TB of bandwidth that doesn’t count against your main VM, plus this content is served off of Vultr’s CDN instead of your instance, meaning even less CPU load for you.
This is pretty easy to do. What we’ll be doing is diverging slightly from the official Lemmy ansible setup to add some different environment variables to pict-rs.
After step 5, before running the ansible playbook, we’re going to modify the ansible template slightly:
cd templates/
cp docker-compose.yml docker-compose.yml.original
Now we’re going to edit the docker-compose.yml with your favourite text editor, personally I like micro but vim, emacs, nano or whatever will do…
favourite-editor docker-compose.yml
Down around line 67 begins the section for pictrs, you’ll notice under the environment section there are a bunch of things that the Lemmy guys predefined. We’re going to add some here to take advantage of the new support for object storage in pict-rs 0.4+:
At the bottom of the environment section we’ll add these new vars:
- PICTRS__STORE__TYPE=object_storage
- PICTRS__STORE__ENDPOINT=Your Object Store Endpoint
- PICTRS__STORE__BUCKET_NAME=Your Bucket Name
- PICTRS__STORE__REGION=Your Bucket Region
- PICTRS__STORE__USE_PATH_STYLE=false
- PICTRS__STORE__ACCESS_KEY=Your Access Key
- PICTRS__STORE__SECRET_KEY=Your Secret Key
So your whole pictrs section looks something like this: https://pastebin.com/X1dP1jew
The actual bucket name, region, access key and secret key will come from your provider. If you’re using Vultr like me then they are under the details after you’ve created your object store, under Overview -> S3 Credentials. On Vultr your endpoint will be something like sjc1.vultrobjects.com, and your region is the domain prefix, so in this case sjc1.
Now you can install as usual. If you have an existing instance already deployed, there is an additional migration command you have to run to move your on-disk images into the object storage.
You’re now good to go and things should pretty much behave like before, except pict-rs will be saving images to your designated cloud/object store, and when serving images it will instead redirect clients to pull directly from the object store, saving you a lot of storage, cpu use and bandwidth, and therefore money.
Hope this helps someone, I am not an expert in either Lemmy administration nor Linux sysadmin stuff, but I can say I’ve done this on my own instance at federate.cc and so far I can’t see any ill effects.
Happy Lemmy-ing!
Great write-up! This might help some of the bigger instances as well.
I assume the larger instances would probably know this already, since they likely have more skilled sysadmin teams than we one-man-show types, but very true - if anything it would save them money to a much greater degree than small instances!
Pretty sure lemmy.world is just a dude, and it’s one of the largest instances.
It’s a team: https://lemmy.world/post/28012
I don’t know much about how they share the load, and Ruud does seem more visibly active than the others… but he’s not a one-man show. In the early weeks he WAS a one-man show, but has a team he works with at mastodon.world and has since brought some of them over to help here.
plus this content is served off of Vultr’s CDN instead of your instance, meaning even less CPU load for you.
Currently the Lemmy backend proxies all image requests, so this isn’t true (for now).
Ah! Noted. That said, it is definitely storing the bytes on the object store. I imagine someone clever with nginx or such could set up some rewrite rules to bypass
pictrsentirely for GET requests, but unfortunately that’s beyond my pay grade here.Not an expert either, but if you do it through nginx I think it will still depend on your single VPS. There probably needs to be a change in the Lemmy-ui to tell the browser to download directly from the object storage CDN.
Thinking about this a little more: I think yeah the HTTP requests will always hit your VPS, but if what you’re saying is that
pictrsis loading from object store and then re-serving them off your VPS, then an NGINX rule might be able to redirect theGETdirectly to the object store; so that instead of transferring the actual image bytes, it just 204’s the browser through to the object store. I don’t know how feasible this is but I may play around with it to see.Pict-rs uses a database to match the uri hash to the file name on disk or object store. This allows for deduplication. It always needs to sit between storage and requests. I have my instance setup to use a separate CDN domain and caching servers to reduce load in my instance. One day soon I hope to get a write done on how to do it.
This allows for deduplication
Really? I’ve found uploading the same image to pict-rs multiple times gives a different hash. It does not seem to dedupe at all.
It allows for different hashes on the front end so individual users can still delete their upload. The sled database maps front end to back end hashes. At least this is what I read from the developer in their matrix chat room.
One would imagine it’s a logical feature, especially for larger images, so maybe it’ll come. For now though, this is still better than not doing it, IMO.
I’d put Wasabi here as well, they’re pretty competitive with their pricing.
Atm I moved to a herzner server woth 500GB SSDs (and consolidated my lemmy insrance eith other stuff I host), and that should be more than enought.
I imagine for larger instances, block storage might xome in handy, even with like running Minio on a storage type server.
Great writeup!
I’m using S3FS to achieve the same thing, but without modifying the ansible config or using native object storage within pict-rs.
deleted by creator
Thank you so much for this. It is much needed!
Oh this is great, thanks!
Can anyone share what bucket permissions they used for pict-rs? I am using minio and used the below policy for an access key, but am still getting unauthorized responses
{ "Version": "2012-10-17", "Statement": [ { "Effect": "Allow", "Principal": { "AWS": [ "*" ] }, "Action": [ "s3:GetBucketLocation", "s3:ListBucket" ], "Resource": [ "arn:aws:s3:::pict-rs" ] }, { "Effect": "Allow", "Principal": { "AWS": [ "*" ] }, "Action": [ "s3:GetObject" ], "Resource": [ "arn:aws:s3:::pict-rs/*" ] } ] }Thank you for this write-up. Your post is the only place I can find on the internet on making the transition to object storage specifically with Lemmy.
Glad you found it helpful!
Thank you for sharing this. I’m going to try to go through this migration shortly.
Right now I’m running my instance on a fairly lean VPS so being able to lighten the CPU load and not have to pre-allocate disk space is super useful.
Replying to confirm that this works and went very smoothly! If you can see my profile picture, it’s on S3 instead of disk now.
I’m using pure ansible to deploy my containers (instead of docker compose) so I had to figure out how to start the pictrs container without actually starting pictrs so that I could run the migration. I ended up stopping the container and then running this to perform the migration:
docker run --name pictrs-migration \ --user 991:991 \ -v /my-pictrs-path/:/mnt \ --rm \ asonix/pictrs:0.4.0-rc.14 \ pict-rs \ migrate-store \ filesystem \ object-storage \ -e https://my-s3-endpoint \ -b my-s3-bucket-name \ -r my-region \ -a my-key-id \ -s my-key-secretThen I used ansible to redeploy the container with volume mount removed and the new s3 environment variables.
Super easy!
I’m attempting this migration on an instance that has been running for about a month, is federated with the top 10+ instances and has synced a lot of data.
The steps I’m using are as follows:
stop docker: sudo docker stop domainname_pictrs_1
run docker-compose to open a session in the stopped container: sudo docker-compose run pictrs sh
run the cmdlet to migrate pictrs via https://git.asonix.dog/asonix/pict-rs/#filesystem-to-object-storage-migration
When this runs, it appers to be trying to sync like… all of the lemmy fediverse… to my object storage:
2023-08-13T17:55:44.426301Z WARN pict_rs: Running checks
2023-08-13T17:55:45.188984Z WARN pict_rs: Checks complete, migrating store
2023-08-13T17:55:45.275403Z WARN pict_rs: 56963 hashes will be migrated
Most of these fail, and I’m trying to run it again with --skip-missing-files , but based on what I’m seeing I don’t know if this is really something that can be done once an instance has federated with a lot of other instances.
Am I missing something?
Edit: with --skip-missing-files its telling me that it’s going to take 23403 seconds (6.5 hours) to complete this migration.
When I look into the bucket, I see all kinds of random images being migrated over, so it’s definitely storing pretty much every image that my instance has ever synced. Is there a way to just migrate content that originated on my instance?
Hi, thanks for lemmy.federate.cc. I will subscribe to any communities there as they come up. Thank you for your service to the fediverse
Awesome!
My current host is mich more expensive. They (Contabo) sell 250GB for €3. Storj.io sell 1TB for €4 for storage, €7 for bandwidth. But it isstille more viable than upgrading my VPS.
You could also consider LPP for purging old posts that your community hasn’t interacted with: https://lemmy.world/post/559690
If using vultr I’d recommend using backblaze for the S3 backend as data transfers will be unlimited. If using vultr object storage data to/from their object storage to the VPS is using the data cap unlike backblaze and vultr who are part of the bandwidth alliance which mean no data usage is counted and if you put your vultr instance behind cloudflare who is also part of the alliance you won’t use any data on the instance for web data, but we’ve seen how cloudflare seems to have caused issue with Kbin so I’m not sure that’s the best thing to do.
I apologize for the run-on sentence :-)
