The paper I linked for Anna is "v0" and has been extended with "v1". The C++ implementation follows "v1". https://arxiv.org/abs/1809.00089 v0 implements any-scale coordination-free partitioned lattice replication with simple multi-master replication using a single number for the whole system. v1 replaces the single number with selective replication (per key), adds vertical tiering (in-memory vs. persistent) and horizontal elasticity (scaling the cluster respectively with load to keep the minimum latency within a bound). This adds two services around the core anna-kvs, namely anna-monitor and anna-router. The monitor watches the cluster and tunes the selective replication such that hot keys are available on many in-memory nodes, while cold keys exist on fewer on-disk nodes. If the network is under or over utilized, the monitor can add or remove resources from the network (if presented the functionality and authority to do so; otherwise it may just warn its administrator accordingly.) The router abstracts away the indirections of the elastic system from clients. v0 was approximately 2k lines of C++ excluding external libraries (zmq and protobufs) but including the lattice library and client code. (Probably excluding comments and blank lines.) v1 as-of-now is 3166 lines of code excluding comments and blank lines. There are a few lattices defined with different consistency semantics. (Causally) (un)ordered (multi)values. I.e. {get,put} key value, {get,put}_set key [values], {get,put}_causal key value. It only takes about 10-20 lines of additional code to implement a different semantic if necessary for your application. Finally, v0 was a prototype and I haven't seen any indication that v1 is "production ready" but it does not appear far from it. I have fixed one issue and have reported a few more. But these are minor and should be easy to resolve. Hint; don't worry about the current 100% CPU utilization... That will likely be fixed by replacing just one async operation with a blocking one.