> New employees will have to learn that instead of using K8S APIs

You know tech has made it when people offer argumentum ad Java.

'Sure, X might not be the best solution, but you can always buy someone to do it.'

It's also very empowering for people whose job isn't to run things, but to build the things that are to run. I've worked on a dozen-or-so of bespoke "industry standard" setups, and each and every one had a number of weird quirks, involved learning some new "industry standard" components and either made it very hard and dangerous for non-ops/devops/infra people to run their things themselves, or had homegrown tooling that pretty much replicated what the k8s API can do, just only a small subset and badly. Some YAML and kubectl are well within what a typical data scientist can be expected to understand, more so if it means they can run their things on a dev cluster themselves, and in a pinch, that data scientist can debug prod issues of their things, because it all works the same way. We have a very useful bot that was built and deployed by someone decidedly not ops while waiting for jobs to finish – a simple K8S deploymend YAML is like 50 LOC, with 40 being pretty much standard, et voilà, running bot, without having to build lots and lots of automation in-house or having to take up ops time to deploy this for them or having to grok advanced sysadmin-ing first. Used with appropriate caution and safeguards, it's super powerful.