pgRx

This project is in early development and highly unstable. APIs, storage format, and wire protocol behavior will change without notice. Do not use in production. Contributions and feedback welcome — just know that everything is actively being reworked.

PostgreSQL-compatible database built on BEAM + Rust.

Same SQL. Same drivers. 100x less memory per connection.

Why

PostgreSQL forks an OS process for every connection — 14.6 MB each, even idle. 50 idle connections cost 729 MB. This architecture hasn't changed since 1986.

pgRx replaces the process-per-connection model with BEAM lightweight processes (~2 KB active, 0 KB hibernated) and a Rust query engine for parsing and execution.

| | PostgreSQL 17 | pgRx | |---|---|---| | Per idle connection | 14.6 MB | 0 KB (hibernated) | | Per active connection | 14.6 MB | ~15 KB | | 50 connections overhead | +729 MB | +768 KB | | Connection model | fork() per conn | BEAM process | | MVCC | Dead tuples + VACUUM | Undo-log (planned) | | Cold start | 0 ms | 0 ms |

Architecture

Client (psql, any PG driver)
  → BEAM process (2 KB per connection, supervised)
    → Rust NIF (parse + execute via pg_query)
      → Shared storage (RwLock, no per-connection duplication)

BEAM handles connections, wire protocol, fault tolerance, and clustering. Rust handles SQL parsing (via pg_query — same parser PostgreSQL uses), expression evaluation, JOINs, and storage.

What Works

Wire Protocol v3 — psql connects without changes
SQL — JOINs (INNER, LEFT, RIGHT, FULL, CROSS), subqueries (IN, EXISTS, scalar, derived), aggregates (COUNT, SUM, AVG, MIN, MAX), GROUP BY, HAVING, ORDER BY, LIMIT, OFFSET
DML — INSERT, UPDATE, DELETE with RETURNING
DDL — CREATE TABLE, DROP TABLE, SERIAL, DEFAULT, constraints (PK, UNIQUE, NOT NULL)
Types — int2/4/8, float4/8, numeric, bool, text, varchar, bytea, serial, vector(N)
Vector — vector(N) type with <-> (L2), <=> (cosine), <#> (inner product) operators — pgvector wire-compatible
Protocol — Simple query, extended query, prepared statements
Connection — Hibernation (0 KB idle), built-in pooling via BEAM
Tests — 97 engine tests passing

Quick Start

# Clone
git clone https://github.com/JagritGumber/pgrx.git
cd pgrx

# Dependencies: Elixir ~1.18, OTP 27, Rust (2024 edition)
mix deps.get

# Build the Rust NIF
cd native/engine && cargo build --release && cd ../..

# Start pgRx
mix run --no-halt

# Connect with psql (default port 5432)
psql -h localhost -p 5432 -U postgres

Roadmap

| Phase | Status | What | |---|---|---| | Wire protocol + basic CRUD | Done | psql connects, INSERT/SELECT/UPDATE/DELETE | | JOINs + aggregates | Done | All JOIN types, GROUP BY, HAVING, ORDER BY | | Subqueries | Done | IN, EXISTS, scalar, derived tables | | Connection hibernation | Done | 0 KB idle connections | | Vector type (pgvector compat) | Done | vector(N), distance operators | | Persistent storage | Next | B-tree indexes, WAL, crash recovery | | ACID transactions | Planned | BEGIN/COMMIT/ROLLBACK, isolation levels | | Undo-log MVCC | Planned | Zero bloat, no VACUUM | | BEAM clustering | Planned | Multi-node distribution |

See FEATURES.md for the complete specification.

Benchmarks

Memory: 50 idle connections

| | PostgreSQL 17 | pgRx | |---|---|---| | Additional memory | +729 MB | +0 KB | | Per connection | 14.6 MB | 0 KB (hibernated) |

Query throughput

pgRx achieves 2x the query throughput of PostgreSQL on equivalent hardware in initial benchmarks.

License

Business Source License 1.1 — the same license used by CockroachDB.

Use pgRx for anything, including commercial production use
Self-host freely
Cannot offer pgRx as a managed database service

Converts to AGPL-3.0 five years after each version's release.

Pgrx

Install / Use

README