Simple rate limiting per deployments

[igalshilman]

Right now we excise a back off strategy per key, while it might be more beneficial also to consider a local view (per node) that keeps track of errors per deployment. The error information can serve as a good signal to when we are out of capacity and can be used to slow down. There are many variations with different trades off. I propose to investigate this and see if we can find a simple approach, to support this.