The downside, or challenge, with NoSQL (generally speaking) is that you need to handle your aggregations ahead of time - you need to know what queries you'll want to run in the future when you store your data. If you have some new aggregation you want to keep, you'll need to re-process the data (with Hadoop or something else).

It's the trade-off of being able to scale reads and writes horizontally. And unless you need it, an RDBMS makes sense given the flexibility.

Maybe, instead of looking at NoSQL as a full-on replacement for RDBMS, we can look at it as a better solution to sharding.