Implement cache invalidation correctly / the best way from the start?

#37 (comment 22350201) @pcarranza pointed something out that crossed my mind as well. The cache invalidation mechanism is a bit of a relay race.

Git client -{HTTP,SSH}-> git receive-pack -(subprocess)-> gitlab-shell/bin/post-receive -(HTTP POST)-> Unicorn -(gRPC)-> Gitaly

Ultimately, I would like that to look like:

Git client -{HTTP,SSH}-> Gitaly -> git receive-pack ; invalidate cache when subprocess finishes

The problem with the latter is that it requires more work up-front:

• forward HTTP POST /receive-pack to gitaly from workhorse
• forward git receive-pack SSH session to gitaly from gitlab-shell

Is having a more robust system early on worth the extra up-front work?

cc @andrewn