Your ML model cache volume is getting blown up during restart and the model is being re-downloaded during the first search post-restart. Either set it to a path somewhere on your storage, or ensure you’re not blowing up the dynamic volume upon restart.
In my case I changed this:
immich-machine-learning:
...
volumes:
- model-cache:/cache
To that:
immich-machine-learning:
...
volumes:
- ./cache:/cache
I no longer have to wait uncomfortably long when I’m trying to show off Smart Search to a friend, or just need a meme pronto.
That’ll be all.
You must log in or # to comment.
It’s not normal for
- model-cache:/cacheto be deleted on restart or even upgrade. You shouldn’t need to do this.Yes, it depends on how you’re managing the service. If you’re using one of the common patterns via systemd, you may be cleaning up everything, including old volumes, like I do.
E: Also if you have any sort of lazy prune op running on a timer, it could blow it up at some point.
Doing a volume like the default Immich docker-compose uses should work fine, even through restarts. I’m not sure why your setup is blowing up the volume.
Normally volumes are only removed if there is no running container associated with it, and you manually run
docker volume pruneBecause I clean everything up that’s not explicitly on disk on restart:
[Unit] Description=Immich in Docker After=docker.service Requires=docker.service [Service] TimeoutStartSec=0 WorkingDirectory=/opt/immich-docker ExecStartPre=-/usr/bin/docker compose kill --remove-orphans ExecStartPre=-/usr/bin/docker compose down --remove-orphans ExecStartPre=-/usr/bin/docker compose rm -f -s -v ExecStartPre=-/usr/bin/docker compose pull ExecStart=/usr/bin/docker compose up Restart=always RestartSec=30 [Install] WantedBy=multi-user.targetWow, you pull new images every time you boot up? Coming from a mindset of having rock solid stability, this scares me. You’re living your life on the edge my friend. I wish I could do that.
I use a fixed tag. 😂 It’s more a simple way to update. Change the tag in SaltStack, apply config, service is restarted, new tag is pulled. If the tag doesn’t change, the pull is a noop.
Ahh, calmed me down. Never thought of doing anything like you’re doing it here, but I do like it.
But why?
why not just down up normally and have a cleanup job on a schedule to get rid of any orphans?
But why?
I a world where we can’t really be sure what’s in an upgrade, a super-clean start that burns any ephemeral data is about the best way to ensure a consistent start.
And consistency gives reliability, as much as we can get without validation (validation is “compare to what’s correct”, but consistency is “try to repeat whatever it was”).
Oh, and if you haven’t changed from the default ML model, please do. The results are phenomenal. The default is nice but only really needed on really low power hardware. If you have a notebook/desktop class CPU and/or GPU with 6GB+ of RAM, you should try a larger model. I used the best model they have and it consumes around 4GB VRAM.
Which model would you recommend? I just switched from ViT-B/32 to ViT-SO400M-16-SigLIP2-384__webli since it seemed to be the most popular.
I switched to the same model. It’s absolutely spectacular. The only extra thing I did was to increase the concurrent job count for Smart Search and to give the model access to my GPU which sped up the initial scan at least an order of magnitude.
Seems to work really well. I can do obscure searches like Outer Wilds and it will pull up pictures I took from my phone of random gameplay movements, so it’s not doing any filename or metadata cheating there.
Is this something that would be recommended if self-hosting off a Synology 920+ NAS?
My NAS does have extra ram to spare because I upgraded it, and has NVME cache 🤗
That’s a Celeron right? I’d try a better AI model. Check this page for the list. You could try the heaviest one. It’ll take a long time to process your library but inference is faster. I don’t know how much faster it is. Maybe it would be fast enough to be usable. If not usable, choose a lighter model. There’s execution times in the table that I assume tell us how heavy the models are. Once you change a model, you have to let it rescan the library.
That’s a Celeron right?
Yup, the Intel J4125 Celeron 4-Core CPU, 2.0-2.7Ghz.
I switched to the ViT-SO400M-16-SigLIP2-384__webli model, same as what you use. I don’t worry about processing time, but it looks like a more capable model, and I really only use immich for contextual search anyway, so that might be a nice upgrade.
What’s your consideration for choosing this one? I would have thought ViT-B-16-SigLIP2__webli to be slightly more accurate, with faster response and all that while keeping a slightly less RAM consumption (1.4GB less I think).
Seemed to be the most popular. LOL The smart search job hasn’t been running for long, so I’ll check that other one out and see how it compares. If it looks better, I can easily use that.
Did you run the Smart Search job?
Running now.
Let me know how inference goes. I might recommend that to a friend with a similar CPU.
I decided on the ViT-B-16-SigLIP2__webli model, so switched to that last night. I also needed to update my server to the latest version of Immich, so a new smart search job was run late last night.
Out of 140,000+ photos/videos, it’s down to 104,000 and I have it set to 6 concurrent tasks.
I don’t mind it processing for 24h. I believe when I first set immich up, the smart search took many days. I’m still able to use the app and website to navigate and search without any delays.
Ok I should know this by now, but what actually is the current “./“ directory when you use that? Is it the docker daemons start dir like /var/docker ?
./will be the directory you run your compose fromI’m almost sure that
./is the directory of the compose.yaml.Normally I just run
docker compose up -din the project directory, but I could rundocker compose up -d -f /somewhere/compose.yamlfrom another directory, and then the./would be/somewhere, and not the directory where I started the command.That seems like a bad idea
As other stated it’s not a bad way of managing volumes. In my scenario I store all volumes in a
/configfolder.For example on my SearXNG instance I have a volume like such:
services: searxng: … volumes: - ./config/searx:/etc/searxng:rwThis makes the files for SearXNG two folders away. I also store these in the
/home/YourUserdirectory so docker avoids using sudoers access whenever possible.So why would you not write out the full path? I frequently rerun compose commands from various places, if I’m troubleshooting an issue.
So why would you not write out the full path?
The other day my raspberry pi decided it didn’t want to boot up, I guess it didn’t like being hosted on an SD card anymore, so I backed up my
composefolder and reinstalled Rasp Pi OS under a different username than my last install.If I specified the full path on every container it would be annoying to have to redo them if I decided I want to move to another directory/drive or change my username.
I’d just do it with a simple search and replace. Have done. I feel like relative paths leave too much room for human error.
Its convenient because your data is stored in the same folder that your docker-compose.yaml file is in, making backups or migrations simpler.
Yup. Everything is in one place and there’s no hardcoded paths outside of the work dir making it trivial to move across storage or even machines.
It’s not.
It usually is the directory where you execute the docker compose command.
