builder: Rewrite in Go to be more reliable and more testable

[carmstrong]

The builder component is notoriously fragile, and has many associated issues, including:

• git push returns 0 on docker error #3317
• Deis apps:destroy does not remove the git repo with the same name #3249
• git push errors could be more helpful than "permission denied" #1772
• config value escaping #2980
• Setting BUILDPACK_URL errs with fatal: I don't handle protocol '"https' #3078

Part of this is due to the fragility of the scripts themselves, but it's also because we cannot write unit tests for the builder component. We need to prioritize rewriting the component in Go so it's something that can be more easily maintained moving forward.

[bacongobbler]

+1

[bacongobbler]

#3684 helps a bit with this. To be clear, I think that the core part of builder (the script that is run when a commit is received) would definitely benefit from a re-factor. However, gitreceive or re-implementing a git server in Go is fruitless. openssh/gitreceive seems to handle that quite well.

[carmstrong]

However, gitreceive or re-implementing a git server in Go is fruitless.

I think that's perfectly reasonable. In my mind, the goals of a builder refactor are twofold:

• Make the codebase less fragile and easier to maintain (many bug fix attempts have resulted in regressions)
• Add test coverage to gain confidence in changes

Having the builder in Go is secondary to both of these primary motivations.

Both of those can be achieved without a new git server. 👍

[carmstrong]

closing in favor of #3758