This article has several misconceptions ignores a few other considerations. consul is doing raft consensus for data replication, and its k/v api is not gossip based, its strongly consistent. the author seems to have missed the comparison to serf on the consul website which might have elucidated some misconceptions that ~'gossip is used to transmit all data'. the article also seems to be under the misconception that you have to run serf binaries as part of consul (its embedded gossip lib). it also ignores the operational simplicity of a single binary vs. managing a zk cluster + smartstack setup. zk (at least till whenever 3.5 comes out) doesn't tolerate dynamic cluster reconfig nicely without round robin restarts. it feels like a pretty biased or uninformed comparison coming from the author of one of the things being compared (partially forgiven as a result of it being a response to a comparison being made by the other software).