I'm the author of a distributed database, competing with them. Overall I'm refreshingly impressed with their design document. Which I am sad to say that most other databases don't come any where near thinking these things through except as an after thought - so I am glad to see they are making it their priority.<p>With that said, they seem to be assuming that their clock skew (ε) has a fixed maximum boundary which is incredibly disconcerting to me as it implies that in certain (rare and anomalous) network partitions that they'll get data corruption and fail.<p>I can see how they, coming from a Spanner background with atomic clocks, might assume this. But this assumption requires that their database cluster is always connected, within some heartbeat interval (which they mention) such that they can trust there exists a maximum bounded ε skew.<p>So while it seems like a dumb question, I honestly must ask a very trivial question: how does CockroachDB handle basic network partitions? I assume they have a good answer to this, but it needs to be clarified in order to answer the more important issue of anomalous partitions, like split brain. This might rip the crockroach in half, quite literally, meaning that all other "guarantees" they give get thrown out the window like linearizability and global consistency.