Installed it because of this thread to my homelab today. I never really managed my phone images in any way, never uploaded them anywhere. This was the first time. About 5 gigabytes of images and videos were synced to my NAS in a few minutes, now I can search them and all that. It's a pretty cool setup, although the installation is a bit tricky if you don't go to the path they give you. I run a Postgres server in Proxmox, and you have to install just the right version of pgvecto.rs for the system to work.
Browsing the issues I was able to figure out what went wrong, and after downgrading, no issues.
But it's not that difficult to dedicate Docker compose file for an "immich project" and use exactly as developer suggests. You are not like going to have 100+ users, more like 1-10 users and even RPI would be enough. It's not an issue to have small database along immich project on the same host.
It just doesn't feel right to have multiple postgres databases running, if every other service uses the one in the network. Having already monitoring, disk space and backups set...
Think this way: postgress db is just part of immich. That's it - separate your services into logical units.
That's actually makes more sense to do at home lab. Bringing down your main DB breaks a lot of your services. By separating - only part would be broken.
My postgress db lives in the same docker compose file where immich is. If I decide to delete immich - it's very simple to run "docker compose down" and delete folders. :)
Ask yourself what your "job" in the homelab should be: do you want to manage what apps are available or do you want to be a DB admin? Because if you are sharing DB-containers between multiple applications, then you've basically signed up to checking the release notes of each release of each involved app closely to check for changes like this.
Treating "immich+postgres+redis+..." as a single unit that you deploy and upgrade together makes everything simpler at the (probably small) cost of requiring some more resources. But even on a 4GB-ram RPi that's unlikely to become the primary issue soon.
Though if you have something better than a Pi that would be ideal. Then it can do nice things like face detection, object detection for search, and transcoding.