So while working on Marmot (https://github.com/maxpert/marmot) I considered something similar, and the thing that stood obvious with my benchmarks was readers can go really fast if I let them do things in process vs over some protocol (be HTTP or Postgres). I had the same problem with rqlite and dqlite, so I made a very conscious design choice to build Marmot into a side-car and not a layer on top. This idea is so tempting and goes against ethos of SQLite IMO.