Make kubeadm job friendlier for local execution.

[pipejakob]

This lets users gracefully invoke the job's Docker image with mounted volumes for the necessary repositories (instead of cloning them).

Previously, they would have to comment out these git clone lines and rebuild the image or encounter an error because the directories already existed.

[fejta]

The fact that this image clones test infra seems to be the core problem.

[pipejakob]

Instead of just closing the PR without merging, can we discuss the best way forward for this?

Cloning the test-infra in order to call bootstrap.py is what almost all of our other images do:

https://github.com/kubernetes/test-infra/blob/master/images/e2e-prow/runner
https://github.com/kubernetes/test-infra/blob/master/images/gcloud/runner
https://github.com/kubernetes/test-infra/blob/master/images/pull-test-infra-gubernator/runner
https://github.com/kubernetes/test-infra/blob/master/images/pull_kubernetes_bazel/runner

Are you suggesting that I bake the test-infra repo into the image instead? Is this PR not an incremental improvement over what's there?

[fejta]

I now understand better why this image exists, but I still do not like its new dynamic behavior. Wouldn't it be easier and more efficient to just run the bootstrap command directly?
/assign

[pipejakob]

You're right that I could run bootstrap.py directly for local execution, but what I'm trying to do is have local execution match production execution much more closely, which means running the entire job through the real Docker image. You might remember that the first iteration of this job in Jenkins required a lot of post-merge debugging precisely because my local testing environment didn't match the way Jenkins ran the job, which led to a lot of small fixes and long feedback loops to see if the job would get any further in its execution before hitting another error. In order to speed up development time, I want fewer surprises between local development and post-merge behavior, which means fewer differences between the two. This change makes it much easier to run the full e2e test locally from within the actual Docker container that will be used in production, but I am open to any other solutions that achieve the same result. This basically codifies what I had been doing locally (commenting out these git clone lines so I could mount local volumes instead), but if there's a better strategy to have more accurate test runs locally, I'm definitely open minded.

[fejta]

Ideally I would like to identify why your local run of bootstrap did not accurate represent what the container would do and we fix that.

However this is good for now.

/lgtm