Support seamless deployments

[grymmy]

Right now, the process of deploying new code to hack.af causes an outage. Hack.af should support seamless rolling deploys such that no traffic is dropped during the upgrade. Here is how I suggest proving we properly support this:

1. Build a script that generates a large amount of hack.af traffic, and outputs a histogram of error codes/latencies
2. Run that script and point it at the hack.af service
3. Deploy new code to hack.af
4. Observe low errors and no detriment to latencies