IME SQL and relational databases have the framework problem, where they expect to be used in a particular way and won't let you use the internals. It's like when a language (ADA?) famously built a high-level version of concurrency into the language, where you were meant to use their "rendezvous" (I've probably mistaken the same), and instead what happened is that programmers implemented mutexes on top of this "rendezvous" and then reimplemented high-level concurrency on top of that.

What we need is databases that expose more of their internals, that are designed to be embedded in applications and used as libraries. All serious databases use MVCC these days, but none of them expose it to the user. All serious databases separate updating the data from updating the index, but few of them expose that to the user. All serious databases know the difference between an indexed join and a table scan, but good luck figuring it out by just looking at an SQL query. Etc.