Programming at Light Speed: Building the Gravity Query Engine

Gravity is the JIT query engine introduced in pgrust v0.4-beta, built to turn analytical SQL into machine code close to the hardware’s practical speed limit. Rather than incrementally tuning an existing executor, its developers first established a ceiling: hand-optimized, multithreaded Rust programs running directly over pgrust’s columnar storage. Those programs ran about four times faster than existing analytical databases, becoming a performance target for the engine. Gravity’s runtime compiler aims to come within 20% of that baseline while retaining Postgres’s SQL dialect and type system.

Compilation proceeds from a standard Postgres plan through a pipeline graph and a Lisp-like intermediate representation to target assembly. Pipelines fuse scans, filters, projections, and aggregation, while the IR lets the compiler apply query-specific optimizations such as inlining, late materialization, and register allocation. The engine compiles queries in under a millisecond and, across analytical workloads, runs within 10-20% of the hand-written Rust baseline. On ClickBench it reaches up to three times ClickHouse’s speed; against pgrust’s Volcano row executor on the same data, it is 102 times faster cold and 499 times faster hot. Because optimizations can fail on less obvious query forms or types, the team catalogued them and built a fuzzer to check that they apply across varied query shapes and Postgres types.