Skip to content

GOAL.md

Latest commit

 

History

History
57 lines (46 loc) · 2.72 KB

GOAL.md

File metadata and controls

57 lines (46 loc) · 2.72 KB
Raw

Goal

Build a from-scratch, idiomatic-Rust rewrite of PostgreSQL 18.3 that is measurably faster than C Postgres — not a port that merely matches it, and not a compatible-but-different database. Same wire protocol, same SQL semantics, same on-disk format (a C 18.3 binary can boot our data directory and vice versa), byte-identical error codes and messages — but faster.

What "faster" means

Every performance claim passes a dual gate on the target hardware (AWS Graviton 4 / Neoverse V2, the fleet's c8gd instances):

  1. ≤ 1.0× C instructions — instruction count at most that of C Postgres built with clang-16 -O3 -mcpu=neoverse-v2, and
  2. ≤ 1.0× C wall time — nanoseconds at most C's on the same machine, same run.

Both, not either: wall-time wins that ride on instruction bloat are fragile across microarchitectures, and instruction wins that lose wall time are not wins. Rust side is pinned rustc, fat LTO, codegen-units=1, -Ctarget-cpu=neoverse-v2.

Claims are made at three levels, each with its own evidence:

  • Per-unit — every hot ported unit gets a microbenchmark against a verbatim C lift of the same code, run on the fleet, with an asm diff when the numbers disagree. Records live in docs/benchmarks/.
  • Composed gates — end-to-end workloads against real C Postgres in the same pod: M1 (SELECT 1 — full protocol/parse/plan/execute round trip), M2 (indexed point selects), M3 (aggregates, sorts, joins).
  • Correctness floor — a differential corpus replayed against live PostgreSQL must stay byte-identical, and C↔pgrust interop on a shared data directory (boot, crash recovery, vacuum WAL) must hold.

How

  • Port PostgreSQL unit by unit from the C source, C-exact in semantics, tracked in CATALOG.tsv (~970 units). Hot query path first; periphery last or never.
  • Idiomatic Rust with day-one performance architecture: bump arenas for no-drop node trees, enum dispatch instead of function pointers, resolve-once carriers, thread-per-backend instead of process-per-backend. Deviations from C's shape are taken only when they are the reason we can be faster, and are proven by benchmark.
  • Anything unported fails loudly with a named panic — no silent wrong answers, ever. The frontier of named panics is the work queue.
  • When a unit loses to C, the loss is attributed at the assembly level and either fixed or recorded with a named lever. No unexplained regressions ride along.

Status shorthand

First official M1 verdict: wall time 0.998× — faster than C — with a 1.22× instruction gap under active attribution. Roughly half the server's code (by C source lines) is ported; SELECT 1 through DML, joins, aggregates, DDL, and VACUUM run end-to-end.