goldenflow-native

Optional native (Rust/PyO3) acceleration kernels for GoldenFlow. The sibling of goldenmatch-native; same abi3 / maturin / Arrow-zero-copy posture.

GoldenFlow stays a pure-Python wheel. pip install goldenflow[native] pulls this compiled runtime, and goldenflow.core._native_loader discovers it (or falls back to pure Python when it's absent). It is off by default at runtime even when installed — see Gating.

What it accelerates

The international phone family. GoldenFlow's pure-Python transforms already resolve the common (NANP) case with vectorized Polars expressions; these kernels accelerate the residual — international numbers that would otherwise hit the phonenumbers library one row at a time. Kernels:

Each takes an Arrow Utf8/LargeUtf8 array (zero-copy in), returns an Arrow array (zero-copy out), and releases the GIL around the compute loop. Any row it can't parse comes back null so the Python reference settles it — the native path is never worse than pure Python.

Dates are intentionally not a native kernel: the Polars str.to_date coalesce fast path already resolves the 4-digit-year formats in vectorized Rust, and a per-row chrono parser would be slower while reintroducing a 2-digit-year parity hazard. The ragged date tail stays on dateutil.

Gating: NANP-only (parity-safe, on by default)

phone is in goldenflow.core._native_loader._GATED_ON, so it runs under GOLDENFLOW_NATIVE=auto (the default) — but only in a provably parity-safe mode.

Characterization across 20 country metadata sets showed the Rust phonenumber port is byte-identical to the Python phonenumbers library except when a +CC international number is parsed with a mismatched default region (the transforms pass "US") and its national number starts with 1 (e.g. +33142685300 → native +3342685300 — the port mis-applies US national-prefix stripping). It also diverges on ambiguous leading-1 inputs (1234567890 → native +1234567890 vs python +11234567890).

Two gates make native authoritative only where it's proven correct:

1. Kernel nanp_only mode — emits a result only for country-code-1 numbers; everything else returns null and falls back to Python.
2. Canonical-NANP acceptance (in transforms/_native.py) — keeps only ^\+1[2-9]\d{9}$ outputs, dropping the malformed 9-digit results the leading-1 ambiguity produces.

So native resolves the canonical-NANP residual the Polars fast path can't reach (alpha like 1-800-FLOWERS, extensions, +1-prefixed forms) ~4-5× faster, and everything international/ambiguous defers to Python. phone_country_code is also gated (country code agrees on all NANP); phone_national/phone_validate stay pure Python (no cheap canonical check). Verified byte-identical to phonenumbers over a 60k mixed corpus in packages/python/goldenflow/tests/transforms/test_native_parity.py.

GOLDENFLOW_NATIVE=0 forces pure Python; =1 enables native for all components (bench/parity lane, no nanp_only restriction — will diverge on intl).

Build

# In-tree dev build (drops goldenflow/_native.abi3.so):
python packages/python/goldenflow/scripts/build_native.py

# Distributable abi3 wheel:
cd packages/rust/extensions/native-flow && maturin build --release