Obviously I'm not in their place, but in hindsight, my first thought when reading their initial issue was "Why not make a container that serves the minimal subset of Rubygems stuff necessary to bootstrap Rubygems, and use that as a source in the event that the main side died and needed to be kickstarted again.

Docker itself used a similar strategy for their build process: the recommended build process uses containers to build the Docker binary, so they made a precompiled binary available that was crafted "the hard way" that could then be used for running the saner build process.

The point that they want to track the versions of gems rubygems.org depends on in a VCS and track the dependency for each version of rubygems.org. You can build a container that includes those, but then you'd have to keep the container in a VCS - that's obviously possible but tangential to the problem.

It would be really helpful to go into some detail on the steps taken to clean up the git history.

It isn't what they did, but my recommendation for most such use cases is the BFG Repo-Cleaner: https://rtyley.github.io/bfg-repo-cleaner/

Don't be deceived by the "java" invocations, it's written in Scala, so it's HN-approved!

Removing directories is simple.

http://git-scm.com/docs/git-filter-branch

But you can do almost anything via custom scripts.

The-other-good-DVCS-that-everyone-almost-forgot-about, Mercurial, has exactly two solutions to the problem of overly large repos which, I presume, contain a lot of BLOBs.

First is Streaming Clones[1], whereby Mercurial server just takes all the files from the repository and transmits them over the transport connection, significantly decreasing TTFB.

The second is Largefiles extension[2], which turns Mercurial into more of a CVCS, but allows for efficient storage and retrieval of large binary files.

[1]: https://hglabhq.com/blog/2014/6/20/working-with-mercurial-ov... [2]: http://mercurial.selenic.com/wiki/LargefilesExtension

I find it funny that the problem was they didn't think enough people know about the `--depth 1` flag to git clone, so the solution is to use submodules (which far fewer people know how to use properly).

You can check for a missing submodule (folder) in your setup script and print helpful advice, but you can't do the same for the large repo: the damage is already done by the time you get to print instructions. In this case things are even easier: the submodule is just a backup for the thing that bundler will do anyways when running "bundle install".

Also, don't forget that "everyone must rebase"...