[nickjackson]

This is super sweet.

It would be great to tie this into a release pipeline, where the release process is actively keeping an eye on failure rates of that service, so that bad deploys could be halted or rolled back automatically.

I was thinking this could work really well when using production integration tests. A percentage of that traffic can be dynamically routed to the newly running services, allowing the release pipeline to ensure the service is functioning correctly before routing any real users.