Databases can decide up-front how prescriptive they want to be about specifying a schema for the data.
When NoSQL databases came to the fore a decade ago, a key point was that they didn't require a schema. The schema could be encoded and enforced in the application rather than in the database. It was thought that schemas were a hindrance in dealing with all of the semi structured data that was getting produced in modern enterprise. So because the early NoSQL systems didn't have a type system, they didn't enforce the standard that all rows in the table have the same structure, they didn't enforce a whole lot.
However, today, most of these NoSQL databases have acquired an SQL interface. Most of them have acquired a rich type system. One of the reasons for this has been the realization that SQL is widely known and reduces the on-board friction in working with a new database. Getting started is easier with an SQL interface than it is with an obscure key-value API. More importantly, having a type system frees application developers from having to remember how a particular value was encoded and to decode it appropriately.
Hence, Cassandra deprecated the Thrift API and made CQL the default. HBase still doesn't support SQL access natively, but use of HBase is increasingly pivoting towards SQL interfaces over HBase, such as Phoenix.
