SQL has strong roots in relational algebra, this makes it harder to implement but easier to verify. If you use concepts like 'eventual consistency' it is much harder to find out wether everything still acts according too the rules in every possible case. Transactions and isolation levels are also based on a clear set of rules, with the same effect: hard to do it right, but not too hard to check if it's done right.

NoSQL generally means non-relational. Google returns to SQL for access to large data, for several reasons. I don't think SQL and the relational model are the only option when it comes to databases. However, when a model is based on some theoretical foundation, it is easier to prove that certain constraints will always be satisfiied. Does anyone has already encountered a candidate that rivals the relational model in this aspect?