[marcinzm]

> But when you measure the TCO, is it really?

I'm in the ML space and every small company I try to avoid EKS. Then I hate my life. Sagemaker, for example, is a giant abstracted away mess with random holes (ie: these types of jobs don't work on this GPU type, etc.) compared to just running things on EKS. The same goes to trying to deploy a more complex third party application. I could just deploy their Helm chart or I could spend a lot of time deploying it somehow in our environment.

[nostrebored]

I had to sell Sagemaker and I agree. It is the wrong abstraction layer without the right escape hatches.

I am super pro Ray for handling these types of workloads now. Huge shout out to anyone here working on maintaining that project.