Large-Repo Benchmarks

Graph-sitter's Rust backend is being validated against pinned large repositories before it becomes the default backend. The current benchmark evidence is for the rust-rewrite branch and branch-built wheels, not a final PyPI release.

Use these numbers as release evidence for the current implementation shape: Python remains the authoring shell, while Rust owns compact parse/index graph storage for the supported subset. Hardware, cache state, and package artifact shape can change absolute times.

Pinned Repositories

Repository	Ref	Commit	Language mode
Apache Airflow	`2.10.5`	`b93c3db6b1641b0840bd15ac7d05bc58ff2cccbf`	Python
Next.js	`v15.0.0`	`51bfe3c1863b191f4b039bc230e8ed5c57b0baf3`	TypeScript, JavaScript, React

Codebase Construction

These measurements use real Codebase(...) construction with CodebaseConfig(graph_backend="rust", rust_fallback="error"). In strict Rust mode, the old eager Python graph is blocked after the compact Rust index builds.

Repository	Python wall	Python max RSS	Rust wall	Rust max RSS	Wall improvement	RSS improvement
Apache Airflow `2.10.5`	18.940s	3469.5 MB	4.085s	266.2 MB	4.637x	13.031x
Next.js `v15.0.0`	24.959s	3100.1 MB	10.771s	435.2 MB	2.317x	7.123x

The Airflow row exercises compact Python files, symbols, imports, import resolution, references, dependencies, and Python compatibility handles. The Next.js row exercises TypeScript/JavaScript files, symbols, imports, exports, relative and tsconfig path resolution, JSX component tag references, dependencies, subclass edges, read-only function call records, and read-only Promise-chain records.

Installed-Wheel uvx Proof

Branch-built wheels are also tested through uvx --from dist/<wheel>.whl graph-sitter ..., which proves package contents and CLI entry points rather than relying on an editable checkout.

Repository	Backend	Parse elapsed	`uvx` outer wall	Sampled RSS	Parse improvement	RSS improvement
Apache Airflow `2.10.5`	Rust strict	4.913s	6.064s	487.0 MB	9.818x vs Python parse	11.148x
Apache Airflow `2.10.5`	Python	48.242s	77.649s	5429.3 MB	baseline	baseline
Next.js `v15.0.0`	Rust strict	10.352s	11.508s	537.5 MB	5.598x vs Python parse	8.383x
Next.js `v15.0.0`	Python	57.956s	78.107s	4505.6 MB	baseline	baseline

The installed-wheel Rust parse paths matched the committed compact golden summaries for both repositories.

Codemod Proof

The same branch-built wheel gates run real write-mode transforms on temporary clones:

Repository	Transform proof	Validation
Apache Airflow `2.10.5`	Adds `from typing import Any` and renames `__getattr__` in `airflow/__init__.py`	Only `airflow/__init__.py` changed
Next.js `v15.0.0`	Adds an import, renames `AppRouterAnnouncer`, and updates its importing usage	Only `app-router-announcer.tsx` and `app-router.tsx` changed

These checks prove that the Rust-backed Python shell can run selected codemods without materializing the old Python graph.

Reproduce

Run the fast local gate:

rust-rewrite/tools/check_fast.sh

Run the opt-in large-repo gate:

rust-rewrite/tools/check_pinned_large_repos.sh

Run branch-built wheel proofs:

rust-rewrite/tools/check_wheel_rust_backend.sh
uv run python rust-rewrite/tools/check_wheel_pinned_python_repo.py --compare-python-backend --run-transform-proof
uv run python rust-rewrite/tools/check_wheel_pinned_typescript_repo.py --compare-python-backend --run-transform-proof

See uvx workflows for the user-facing CLI form.

Caveats

Rust is still opt-in. The Python backend remains the default until rollout gates, parity gates, and release artifact validation pass.
The installed-wheel results are branch-built wheel results. Published-package uvx graph-sitter ... claims require a separate clean-environment release transcript.
Counts are compared against compact Rust golden summaries, selected semantic parity checks, and codemod file-diff assertions. Full graph-wide semantic equality is still tracked separately before default-backend promotion.