Buildx in the automated CI environment

[dee-kryvenko]

I am currently exploring buildx - primarily for parallel bake for multi-flavor images and also a caching. I got a couple of questions - documentations seems to be pretty scarce and I couldn't find any answers...

A bit of context - if I were to run buildx in the automated K8s CI environment with ephemeral build pods - I wouldn't want my build pods to have access to k8s to schedule more pods etc. I would want to have a number of buildx builders running and ready to go. Also obviously CLI is not a way to go - I'll want this configuration in code, so when I do docker buildx build - it would connect to pre-existing farm. So a number of questions:

1. Can I deploy buildx with helm/kustomize/etc? Is it just a deployment or anything else?
2. If I then create ~/.docker/buildx/instances/k8s on my build agent as a pipeline step, will it pick it up without ever running docker buildx create?
3. How can I point ~/.docker/buildx/instances/k8s to a running builders via a service endpoint or a DNS? Something tells me "Endpoint":"kubernetes:///k8s?deployment=\u0026kubeconfig=" uses k8s API to discover endpoints, can it point to a DNS record like a service or a ClusterIP?
4. How do I scale it? It looks like CPU/Memory might be a bad metric for HPA, also my understanding is 1 builder can handle 1 build simultaneously. Is there a prometheus endpoint for scraping so I can configure scaling on a custom metric like a number of concurrent builds?
5. Can CLI<->builder communication be encrypted and authenticated?
6. Or am I completely misunderstanding how this suppose to work?