I’ve had this idea floating around in my head for a while: if you could take all the queries and the schema, you should be able to compile that to a specialized binary that only does those queries. I have no idea how to accomplish that but I bet you could do a lot of nifty optimization. Migrations get hard but I wonder how good it would be
*"DBToaster: Higher-order Delta Processing for Dynamic, Frequently Fresh Views"∗
edit: so this only does the read side right? It doesn't look like this supports mutations
The performance is excellent but modern databases tend to be bandwidth-bound for many workloads anyway, so the benefit will be significantly limited.
There is an argument for this, it's just probably not worth the tradeoff. The security benefit seems somewhat reasonable. Feels like the kind of tradeoff as a unikernel.
Also the SQLite bytecode engine. https://www.sqlite.org/opcode.html