Intriguing post! For those who hadn't heard about DRBD, it's http://en.wikipedia.org/wiki/Distributed_Replicated_Block_De... - it makes failover incredibly simple to integrate by pushing a lot of distributed-systems responsibility down to the filesystem level, but in doing so it forces you to have a single-master setup, and you can't take advantage of parts of your domain model that are well-suited to eventual consistency. Thinking about it, now that Riak supports strong (quorum-based) consistency on a per-bucket basis, Riak-backed scalable Git hosting would probably be relatively easy to implement. Looks like you were building your own system - did you ever release the code?
The git object data itself goes into a blob-store (like S3); it can be stored without strong consistency. It turns out you only need to keep track of a very small amount of metadata consistently (the refs). Riak, etcd, DynamoDB or the Google Cloud DataStore would all be good choices, I think.
I was working on an open-source implementation of Raft as part of this (called barge), but it isn't as reliable as the alternatives above - yet!
