scripts/create_safe_public_subset.py

#!/usr/bin/env python3
from pathlib import Path
import argparse
import json
import shutil


def fail(msg):
    raise SystemExit(msg)


def import_arrow():
    try:
        import pyarrow as pa
        import pyarrow.parquet as pq
        import pyarrow.compute as pc
        return pa, pq, pc
    except Exception as e:
        fail(f"pyarrow is required for safe subset creation: {e}")


def auto_col(cols, candidates):
    lower = {c.lower(): c for c in cols}
    for cand in candidates:
        if cand.lower() in lower:
            return lower[cand.lower()]
    return None


def is_lfs_pointer(path):
    try:
        return path.read_bytes()[:120].startswith(b"version https://git-lfs.github.com/spec/v1")
    except OSError:
        return False


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--root", default=".")
    parser.add_argument("--out", default="safe_public_subset")
    parser.add_argument("--risk-column", default=None)
    parser.add_argument("--id-column", default=None)
    parser.add_argument("--exclude-risk-values", default="high", help="Comma-separated risk values to exclude")
    parser.add_argument("--overwrite", action="store_true")
    args = parser.parse_args()

    pa, pq, pc = import_arrow()

    root = Path(args.root)
    out = Path(args.out)

    persona_path = root / "persona_core.parquet"
    if not persona_path.exists():
        fail("persona_core.parquet not found")
    if is_lfs_pointer(persona_path):
        fail("persona_core.parquet is a Git LFS pointer. Pull the real Parquet payload before creating a safe subset.")

    table = pq.read_table(persona_path)
    cols = table.column_names

    if out.exists():
        if not args.overwrite:
            fail(f"Output directory exists: {out}. Use --overwrite to replace it.")
        shutil.rmtree(out)
    out.mkdir(parents=True)

    risk_col = args.risk_column or auto_col(cols, [
        "disclosure_risk_level",
        "disclosure_risk",
        "privacy_risk_level",
        "risk_level",
    ])
    if not risk_col:
        fail(f"Could not auto-detect disclosure risk column. Columns: {cols}")

    id_col = args.id_column or auto_col(cols, [
        "synthetic_person_id",
        "persona_id",
        "person_id",
        "synthetic_person_id",
        "id",
    ])
    if not id_col:
        fail(f"Could not auto-detect persona id column. Columns: {cols}")

    exclude_values = [x.strip().lower() for x in args.exclude_risk_values.split(",") if x.strip()]
    risk_lower = pc.utf8_lower(table[risk_col].cast(pa.string()))
    excluded_mask = pc.is_in(risk_lower, value_set=pa.array(exclude_values))
    keep_mask = pc.invert(excluded_mask)
    safe_persona = table.filter(keep_mask)
    high_persona = table.filter(excluded_mask)

    removed_ids = set(high_persona[id_col].to_pylist())
    pq.write_table(safe_persona, out / "persona_core_safe.parquet")

    report = {
        "risk_column": risk_col,
        "id_column": id_col,
        "exclude_risk_values": exclude_values,
        "persona_core_rows_before": table.num_rows,
        "persona_core_rows_after": safe_persona.num_rows,
        "removed_persona_rows": len(removed_ids),
        "artifacts": {
            "persona_core": {
                "input": "persona_core.parquet",
                "output": "persona_core_safe.parquet",
                "rows_before": table.num_rows,
                "rows_after": safe_persona.num_rows,
            }
        },
    }

    for name in ["persona_views", "actor_state_init"]:
        p = root / f"{name}.parquet"
        if not p.exists():
            continue
        if is_lfs_pointer(p):
            report["artifacts"][name] = {"input": p.name, "skipped": "Git LFS pointer; real Parquet payload not present"}
            continue
        t = pq.read_table(p)
        if id_col not in t.column_names:
            report["artifacts"][name] = {"input": p.name, "skipped": f"missing id column {id_col}"}
            continue
        remove_mask = pc.is_in(t[id_col], value_set=pa.array(list(removed_ids), type=t[id_col].type))
        keep = pc.invert(remove_mask)
        safe_t = t.filter(keep)
        out_name = f"{name}_safe.parquet"
        pq.write_table(safe_t, out / out_name)
        report["artifacts"][name] = {
            "input": p.name,
            "output": out_name,
            "rows_before": t.num_rows,
            "rows_after": safe_t.num_rows,
        }

    (out / "safe_subset_report.json").write_text(json.dumps(report, indent=2, ensure_ascii=False), encoding="utf-8")
    print(json.dumps(report, indent=2, ensure_ascii=False))


if __name__ == "__main__":
    main()