rer

r.e.r stands for the French phrase "Rez En Rust" — a pun on the Parisian suburban train, the réseau express régional d'Île-de-France.

What it is

rer is a Rust reimplementation of the solver hotpath of rez, the VFX/animation package manager.

The solver is a faithful port of rez's own phase-based backtracking solver (rez/src/rez/solver.py) — not a different algorithm dressed up to look similar. It reproduces rez's mechanics exactly: weak (~) and conflict (!) requirement semantics, variant selection order, the extract / intersect / reduce / split cycle, and implicit backtracking. The goal is output that matches rez 1:1, not merely "a valid resolve".

The end goal is a hybrid integration: a Rust library callable from Python via PyO3 that accelerates rez resolves while leaving the rest of rez untouched.

Status

• ✅ Solver — complete and rez-faithful.
• ✅ Validated 1:1 — against rez's bundled 188-case benchmark dataset, every resolve matches rez's own recorded result exactly (same status, same package set).
• ✅ Fast — on that benchmark, on one machine, rer resolves all 188 requests in ~44 s versus ~206 s for rez 3.3.0 (rez benchmark). Correctness is version-independent; the timing is same-machine context, not a lab claim.
• ✅ Python bridge — pyrer.solve(...) runs the ported solver (the rer-python crate ships to PyPI as pyrer — rer is taken; pip install pyrer, import pyrer).

Workspace

A virtual Cargo workspace — use -p <crate> for crate-specific commands.

rer works on an in-memory package repository (family → version → {requires, variants}); it does not read the filesystem itself — the host (rez) hands the loaded data in.

Using it from Rust

use std::rc::Rc;
use rer_resolver::rez_solver::{Requirement, Solver, SolverStatus};
use rer_resolver::PackageData;

let mut repo = std::collections::HashMap::new();
// app-1.0.0 requires lib-2; lib has 1.0.0 and 2.0.0
repo.insert("app".into(), [("1.0.0".to_string(), PackageData {
    requires: vec!["lib-2".into()], variants: vec![],
})].into_iter().collect());
repo.insert("lib".into(), [
    ("1.0.0".to_string(), PackageData::default()),
    ("2.0.0".to_string(), PackageData::default()),
].into_iter().collect());

let reqs = vec![Requirement::parse("app")];
let mut solver = Solver::new(reqs, Rc::new(repo)).unwrap();
solver.solve();
assert_eq!(solver.status(), SolverStatus::Solved);
// resolves to app-1.0.0 + lib-2.0.0

Using it from Python

python -m venv .venv && . .venv/bin/activate
pip install maturin
cd crates/rer-python && maturin develop

import json, pyrer

repo = {
    "app": {"1.0.0": {"requires": ["lib-2"], "variants": []}},
    "lib": {"1.0.0": {"requires": [], "variants": []},
            "2.0.0": {"requires": [], "variants": []}},
}
result = pyrer.solve(["app"], json.dumps(repo))
print(result.status)    # "solved"
print(result.resolved)  # [("app", "1.0.0", None), ("lib", "2.0.0", None)]

solve() reports failures and bad input via result.status ("solved" / "failed" / "error"), never as a Python exception.

Building & testing

cargo build                          # build all crates
cargo test                           # unit + integration tests
cargo bench                          # benchmarks (rer-version)

The rez benchmark

The 1:1 differential test against rez's bundled benchmark is #[ignore]d (the full release run takes several minutes). It needs the rez git submodule:

git submodule update --init
python scripts/prepare_benchmark_data.py     # -> data_set/benchmark_*.json
cargo test --release -p rer-resolver --test test_rez_benchmark -- --ignored
cargo run  --release -p examples --example rez_benchmark_dataset   # timing report

The benchmark CI workflow runs this on every PR touching the resolver.

License

See LICENSE.