https://github.com/bup/bup strategically breaks files to chunks of ~8k length, in a way that makes small changes to a big file become a small change to a small number of chunks (ideally, changing one byte in a 80GB file would change exactly one such 8KB chunk -- and that is often the case, though it's common than 3 or 4 would change).

bup then puts[0] the chunks into git, together with a "reconstruction map". As a result, you can efficiently put huge files in git and only pay for the real deltas in storage.

[0] It doesn't actually use git or libgit - it writes git packs directly.

I really like git annex since the repository is still completely git compatible.

I agree git-annex is the best solution for now - I just wish some effort would go into versioning binary files by creating projects that mimic the behavior of the programs that generate them in some way.

Probably the closest thing to what I am optimistically hoping to see someday is how developers save DB migration files instead of versioning copies of the database tables, and then run the migrations as part of the merge process to sync the databases.

Thinking about it more, it seems like a general project to create 'migration' specifications to generate binary files would be more appropriate than a project to modify version control.

Anyway, thanks for your work on Gitlab!

I agree that it would be nice to have more files being created algorithmically instead of being binary. You can diff an svg image but you can't diff a jpeg. Although the svg diff could be presented better than how we currently do it.

And you're welcome, more than 700 contributors made GitLab what it is today.