Non-local database over high-perf network

[tired echo]

I have a server that can't have GPUs in it but I'm interested in hardware video transcoding and possibly accelerated ML.

I'm thinking about getting a cheap mini-PC with a suitable iGPU and either just using a small boot disk in it or netboot it from the storage server; but in either case run Immich with the data and the database both stored on the storage server and accessed through the network. This is so that I can 1) keep the data safe 2) manage backups in one place 3) it just seems like an interesting project. I can have 10 Gbps connection between the two.

Does this sound feasible and sensible? I think the big problem is Postgres over network, this is likely not a good idea.

[tawdry spire]

Accessing postgres over network is perfectly fine, even on just 1g

Having postgres access its storage over network is where the problems lie

[sinful fjord]

I have done as you ask.
I run Immich in Docker containers over a NAS server without any GPU capability. It took 85 hours to ML-train 85k pictures.
Then I learnt from the docu., that I might duplicate just the ML-container on another computer attached to the same LAN (local network) as the NAS was sitting on.
My old tower PC was rebuilt with new everything in it in 2022. It has 2,5 Gigabits/sec network speed against the LAN, and then also to the NAS.. But no graphics card. But my son-in-law had a used Asus graphics card with a Nvidia TX1070 GPU on it. So, I set up the ML-conatiner on that pc, and told the Immich configuration in the main Immich server of that containers LAN-address.
The result is, that in stead on 85 hours processing time for ML-assets, it now takes only 3 hours 35 minutes. I.e. 23 times faster.

I have set it up, so the ML-container on the pc is prioritised number one, but I have kept the ML-conatiner on the NAS, which then serves a fallback-possibility, if for instance the pc is not up and running, when adding only a ferw new pictures, or doing an intelligent search for specifik pictures.