Comment 2 for bug 894279

I agree that Nova needs to be fixed but we need a cloud-init fix for the current Nova as well. As for running an API server on each compute node, I don't think that's a viable solution for our environment. We've modified memcached to cache the metadata requests and that seems to help make things faster.

I see your point about the metadata service not being ready in time. Do you know how long it can take for the service to become ready? How about a smarter sleeptime that always increases (linearly) from 1 to 20 secs (or some configurable upper limit)?
sleeptime = int(19 * x / (sleeps - 1)) + 1
With that, the 5 retries would happen over a time of 51 secs (1, 5, 10, 15, 20 sec sleeps in between the 5 requests).