Do we need support for "atomic", or "blue green" deployments?

[jbardin]

We currently roll out deployments as fast as we can through the cluster, but we have no way of synchronizing those events, which leads to a mix of versions running for a period of time. This hasn't caused any problems yet, but it seems that at some point an incompatibility between version N and version N+1 will cause a problem.

[joeshaw]

The ability to do an atomic switchover would be nice, but the compatibility issues we tend to have surround the database schema. For those it'd almost be better to queue up and block incoming requests and route them to only new app instances once that migration happens.