Comment 2 for bug 1429087

That might not be a helpful comment, but this issue has rarely been seen in upstream gates and reported by different people.

The problem for sure exists in neutron, but repro or env with a repro not necessarily will help to analyze it. It's because 'lost' request probably doesn't leave a trace in neutron's logs.

So I'd propose the following ways to troubleshoot the issue (if another repro is made):
1) look at haproxy logs for the lost request, try to understand, which neutron server was serving it
2) if backend is identified - try to look at server logs for the request

I guess (2) will not give results.

Another guess is that the issue may have higher rate of occurence if load is increased, for that I'm proposing shutting down neutron-servers on 2 out of 3 controllers.