I've been thinking along similar lines, I think. One problem I see with data in RDBMSs is that data gets transformed in the process of importing it to match the

I've been thinking along similar lines, I think. One problem I see with data in RDBMSs is that data gets transformed in the process of importing it to match the schema, and this transformation loses information, i.e. it's hard to recover the original meaning and context, because it's been contorted to fit the schema. Schemas tend to need to evolve, and so you'd like not to be stuck with the assumptions of schema 1.0 forever. Data isn't that big these days in relation to the speed of computers, so you could keep the data in its original form and leave open the ability to "reimport" to fit the current schema as needed. If you're building a website, for instance, then your web app is going to be doing a limited number of queries over and over. If you start from the queries, you can infer things about how the data should be organized, e.g. which indexes are needed. In fact, if you took a query and pre-generated all the possible results of that query and saved them, that's basically an index, the one that fits that query. The inputs to this hypothetical database would be the original unprocessed data, the set of queries you need, and mappings to transform the unprocessed data into queryable form, i.e. the equivalent of a schema. The inputs apart from the data proper would also be considered as data.