

I see, fair enough. Replication is never instantaneous, so do you have definite bounds on how much latency you’ll accept? Do you really want independent git servers online? Most HA systems have a primary and a failover, so users only see one server. If you want to use Ceph, in practice all servers would be in the same DC. Is that ok?
I think I’d look in one of the many git books out there to see what they say about replication schemes. This sounds like something that must have been done before.
Are you familiar with git hooks? See
https://git-scm.com/book/en/v2/Customizing-Git-Git-Hooks
Scroll to the part about server side hooks. The idea is to automatically propagate updates when you receive them. So git-level replication instead of rsync.