Jepsen showed elasticsearch was pretty bad at partition tolerance. It was a pretty early version, but it would have been nice to see how much they fixed with another Jepsen run.
And distributed SQL of any complexity doesn't really scale. As soon as data that a join depends on is on other machines/nodes, you need quorums and dependent network I/O. Granted with high-speed networks that is getting to be less of a burden speed-wise, but the reliability problems still exist.
