[tryauuum]

god do I hate the "pet" analogy

yes this server is special to me. No, I cannot move workloads somewhere else and shoot the server in the head. Because customers are running GPU workloads there, I cannot transparently migrate them

[fragmede]

How does the analogy falling flat for you? GPUs are interchangable resources; an H100 is just like any other H100, even if you can't take a snapshot of an intermediate state from one and load it onto another and resume.

[tryauuum]

I thought about it and I don't hate the analogy. Although usually I'm very against comparing IT stuff with physical stuff because it tends to produce pointless conversations.

What I hate is the attitude of "treating your server is a pet is wrong, you should be able to kill it and replace any time". I feel like people who say it never run anything more complex than an HTTP API.

[fragmede]

Ah yeah I totally agree. Having pets is fine until it isn't. A single-digit handful is fine. 1,000 bespoke pets, way above Dunbar's number, isn't. Where exactly you run out of mental headroom and should convert is up to you; the point is that there is one and converting over should be planned for in the North Star roadmap of future development.

My (Raspberry Pi) home server is very much a pet, but I'm never going to productionize that thing, so I'm totally fine with it being a pet, despite the practice at work of having no pets, all cattle.