We're having a lot of problems with bamboo deciding our git cache is corrupt, and redownloding it. (It's rather large, approx. 2Gb, which I know is not recommended but I have to live with it for now).

What I think is happening is that while one local agent is checking out source, another agent does a "fetch" and so changes the contents of the repository, meaning the first agent fails. The first agent then decides the cache is corrupt and re-downloads it.

From poking around and seeing what commands bamboo is sending to git, it appears to be initialised as a standard repository - i.e. the command "git init" rather than "git init --bare" - I wonder if this could be the cause?

At the moment most of our agents spend longer updating source to latest than they do building.