Minor Description
Currently, mappings between nodes and collectors change when nodes become disconnected, or when new nodes join the cluster (though through consistent hashing, most existing connections do not wind up getting remapped, preventing the system from thrashing about too much). It would be interesting to allow for custom FailoverChainManager implementations that can adjust mappings based on custom load-based policies, as this would allow one to deal with situations when the collectors are not equivalent (due to bad or busy machines, unexpectedly skewed load, etc). The load could be retrieved from stats in ReportManager.