Don't check in vendor for bootstrap; it adds 160M which slows down cloning the registry as part of Kubeflow deployment

[pdmack]

With the bootstrap vendor dir, a top-level git grep becomes ineffective for some general queries. There seems to be a lot of code here, worthy of its own home IMHO.

[jlewi]

Does it make sense to move bootstrap into its own repo? If we moved the ksonnet registry into its own registry would that help? Or is that just an orthogonal issue?

[pdmack]

# find bootstrap/ -type f | wc -l
9714

[jlewi]

So most of that is just vendor

find ./vendor -type f | wc -l
9692

Is it not possible to simply exclude vendor for git grep?

[pdmack]

git grep is just a pain point. Does bootstrap have references to the rest of the kubeflow src tree?

[jlewi]

No I think it could be moved to its own repo. My only reticence is that it would be unfortunate if we have to create a new repo each time we add a go program because there is a cost to managing each repo.

@kunmingg What do you think?

[jlewi]

The vendor directory is currently 160M. Since our Kubeflow deployment scripts clone the repo a large repository is highly undesirable since it negatively impacts user experience.

Since we use glide to manage dependencies we don't need to check in the dependencies; we could just run glide install.

This comes at the expense of introducing some flakiness into the build process we can mitigate that by

1. Creating a Docker image with glide installed suitable for building go binaries.

2. Adding retries as needed to handle flakes pulling dependencies.

Moving it into its own repository would also work but that comes with its own headaches. I think I'd prefer to just strip out vendor.

[jlewi]

This is fixed.