It seems like there could be a sweet spot of their nice query language, schemaless, and easy scaling for moderate size data sets?
I'm kinda tired of having to go all in on a big hadoop-ecosystem just to figure out average X to Y in a dataset....
Some of our biggest users (to be announced in a few weeks for 2.0) use RethinkDB this way. You can't really do deep analytics/machine learning as RethinkDB wasn't designed for that, but if you want to store a lot of data, and then run lightweight aggregation or map-reduce queries on that data, Rethink turns out to be a really good product for it.
One issue I see with this path is that if your queries ever get a lot more complex, you'd have to migrate off of RethinkDB onto Hadoop (which is a pain). I think that if you know for certain you just want lightweight querying capabilities RethinkDB can be really wonderful, but if there is a good chance you might need something deeper, it might be worth the effort to set up Hadoop early on.
It's not really an MVCC thing, and you can work around it in data model, but for lots of reports (say, running in a cronjob), I want to run a query "as" the database saw things from at midnight UTC, even if i start running it at 2am? It would also make reports a more reproducible... but maybe this is really a datamodel problem. I felt when I read the Google Spanner papers that it was a pretty potentially useful feature for Read-Only queries.