A brief response to "COPS and PNUTS"
I stumbled across a review by Peter Bailis of one of our SOSP papers today, and I'm worried that we failed to be clear about a few things in the work, so I figured I'd grab the blog soapbox and try to clarify. First is the relationship of COPS [ pdf ] and prior work. Peter comments: I'm not sure this advances much of the state of the art beyond, say, Dynamo, which provided a scalable web store with user-specified conflict handling. The delta from Dynamo is very simple: Dynamo offers no “guaranteed” consistency semantics at all. Operations that touch different keys/tablets can arrive at replicas in arbitrary order. In COPS, that’s not the case - operations can only arrive in an order allowed by causal consistency. The example of updating two keys, one that controls the permissions of an object, and one that modifies the object, is a good way to illustrate this. In Dynamo, the object modification could be applied at a remote datacenter before the corresponding permi