A database system that writes its own code while it runs just posted a 59% cost cut on a standard benchmark.
Researchers built KathDB-FAO, a new way for the KathDB database to answer questions asked in plain English. Instead of running a fixed set of prewritten operations, it breaks a query into small, well-defined actions, checks what each one needs and produces, and groups the compatible ones together. Then, for each group, it generates - synthesizes - a custom piece of code on the spot, tailored to that specific query. Tested against other systems on the SemBench benchmark, it cut execution cost by 58.8% on average, without sacrificing answer quality.
Most databases run one of a handful of generic strategies no matter what you ask, which wastes cycles on questions that don't need the full machinery. Writing bespoke code per query - rather than picking from a preset menu - is a bet that generated code can beat hand-tuned code, at least for now. If that holds up outside a single benchmark, it is a meaningful shift in how databases that handle natural-language queries get built.
One benchmark and one lab's numbers are not proof of a trend - and 'writes its own code for every query' is also exactly the kind of complexity that tends to bite back in production.