"""Parallel row-validation harness, shared by every entry. Generalized from chartforge/render_check.py, whose ProcessPoolExecutor driver was always generic while only its inner check was chart-specific. THE RULE THIS MODULE EXISTS TO ENFORCE: **the gate imports the released scorer**, it does not reimplement it. If the scorer's semantics change, the gate changes with it, so a row can never pass generation-time validation and then fail evaluation-time scoring. Every entry passes its own `check_fn`, and every `check_fn` must call into that entry's real scorer rather than a lookalike. A `check_fn` takes whatever tuple the entry finds convenient, whose first element is the row id, and returns `(row_id, ok, reason)`. Reasons are formatted `"bucket: detail"` so `drop_reasons()` aggregates them for free. from common.exec_gate import check_many, drop_reasons verdicts = check_many(items, my_check, workers=8) kept = [i for i in items if verdicts[i[0]][0]] print(drop_reasons(verdicts)) """ from __future__ import annotations import json import subprocess import sys import tempfile from collections import Counter from concurrent.futures import ProcessPoolExecutor, as_completed from pathlib import Path from typing import Callable, Sequence TIMEOUT_S = 8 CheckFn = Callable[[tuple], tuple[str, bool, str]] def check_many(items: Sequence[tuple], check_fn: CheckFn, workers: int = 8, progress_every: int = 2000) -> dict[str, tuple[bool, str]]: """Validate rows in parallel. Execution is CPU-bound, so processes beat threads. check_fn must be a module-level function (picklable), not a lambda or closure. """ results: dict[str, tuple[bool, str]] = {} with ProcessPoolExecutor(max_workers=workers) as pool: futs = {pool.submit(check_fn, it): it[0] for it in items} done = 0 for fut in as_completed(futs): rid, ok, why = fut.result() results[rid] = (ok, why) done += 1 if progress_every and done % progress_every == 0: kept = sum(1 for v in results.values() if v[0]) print(f" checked {done}/{len(items)} kept {kept}", flush=True) return results def drop_reasons(verdicts: dict[str, tuple[bool, str]]) -> dict[str, int]: """Aggregate failures by the bucket prefix before the first colon.""" return dict(Counter(why.split(":")[0] for ok, why in verdicts.values() if not ok ).most_common()) def run_python(code: str, marker: str = "__RESULT__", timeout: int = TIMEOUT_S, preamble: str = "") -> object | None: """Execute code in a subprocess and return the JSON printed after `marker`. Used by entries that score by executing generated code. The subprocess is the isolation boundary: a row that hangs or segfaults costs one worker, not the run. """ script = (preamble + "\n" + code) if preamble else code path = None try: with tempfile.NamedTemporaryFile("w", suffix=".py", delete=False) as f: f.write(script) path = f.name out = subprocess.run([sys.executable, path], capture_output=True, text=True, timeout=timeout) for line in out.stdout.splitlines(): if line.startswith(marker): return json.loads(line[len(marker):]) return None except Exception: return None finally: if path: try: Path(path).unlink(missing_ok=True) except Exception: pass def _echo_check(item: tuple) -> tuple[str, bool, str]: """Module-level so the selftest can pickle it.""" rid, value = item if value < 0: return rid, False, "negative: value below zero" if value == 0: return rid, False, "zero: value is zero" return rid, True, "ok" def _selftest() -> None: items = [("a", 1), ("b", -1), ("c", 0), ("d", 5), ("e", -2)] v = check_many(items, _echo_check, workers=2, progress_every=0) assert len(v) == 5 assert [k for k in v if v[k][0]] == ["a", "d"] or set(k for k in v if v[k][0]) == {"a", "d"} assert drop_reasons(v) == {"negative": 2, "zero": 1}, drop_reasons(v) assert run_python('print("__RESULT__" + __import__("json").dumps([1,2]))') == [1, 2] assert run_python('raise SystemExit(1)') is None assert run_python('import time; time.sleep(30)', timeout=2) is None print("exec_gate selftest: 5/5 OK") if __name__ == "__main__": _selftest()