Berkeley DB, as it's being used in the existing bitcoin client, is a perfect example of why not to use NoSQL.
The type of NoSQL the poster was talking about here appears to be the non-ACID type of which BDB is not one of...
Writing blocks to a flat file is not very ACID, and there is no attempt at something analogous to BEGIN TRANSACTION/COMMIT when updating the block chain or indices, so a failure part-way through a reorg is very likely to leave the database in an inconsistent state. Granted, both of these problems are outside of BDB and do not reflect on the strength or weakness of BDB.
I would say that bitcoin is effectively using a NoSQL "style" by not taking advantage of consistency guarantees, or of a relational model, both of which I think it would benefit from.