"""Feishu webhook monitor for long-running Perturb benchmarks. Watches a benchmark output directory, counts completed samples, estimates ETA, and sends Feishu (Lark) notifications on progress, abnormal exit, or completion. Usage: # Use env var for webhook export FEISHU_WEBHOOK_URL="https://open.feishu.cn/open-apis/bot/v2/hook/xxxx" python scripts/monitor_benchmark_feishu.py \ --output-dir /workspace/Perturb/benchmark_2000_v15 \ --interval-minutes 30 # Or pass webhook directly python scripts/monitor_benchmark_feishu.py \ --output-dir /workspace/Perturb/benchmark_2000_v15 \ --webhook https://open.feishu.cn/open-apis/bot/v2/hook/xxxx """ from __future__ import annotations import argparse import json import os import sys import time from datetime import datetime, timedelta from pathlib import Path import requests def send_feishu(webhook_url: str, text: str) -> bool: """Send a plain text message to a Feishu webhook.""" if not webhook_url or webhook_url.startswith("https://YOUR_"): print(f"[feishu skip] webhook not configured. Message would be:\n{text}") return False payload = { "msg_type": "text", "content": {"text": text}, } try: resp = requests.post(webhook_url, json=payload, timeout=10) resp.raise_for_status() print(f"[feishu ok] {resp.status_code} {resp.text[:120]}") return True except Exception as exc: print(f"[feishu err] {exc}") return False def read_indices(output_dir: Path) -> tuple[int, list[int]] | None: """Return (total, indices) from indices.json if it exists.""" path = output_dir / "indices.json" if not path.exists(): return None try: data = json.loads(path.read_text()) if isinstance(data, list): return len(data), data return int(data.get("n", 0)), data.get("indices", []) except Exception as exc: print(f"[warn] failed to read indices.json: {exc}") return None def count_completed(output_dir: Path, subdir: str) -> int: """Count completed *.json files under output_dir/subdir.""" target = output_dir / subdir if not target.exists(): return 0 return len(list(target.glob("*.json"))) def read_summary(output_dir: Path, subdir: str) -> dict | None: """Read summary.json if available; fall back to aggregating summary_worker*.json.""" target_dir = output_dir / subdir path = target_dir / "summary.json" if path.exists(): try: return json.loads(path.read_text()) except Exception: pass # Multi-worker mode: aggregate per-worker summaries. worker_summaries = [] for ws_path in sorted(target_dir.glob("summary_worker*.json")): try: worker_summaries.append(json.loads(ws_path.read_text())) except Exception: pass if not worker_summaries: return None total_n = sum(ws.get("n", 0) for ws in worker_summaries) if total_n == 0: return None avg_score = sum(ws.get("avg_score", 0.0) * ws.get("n", 0) for ws in worker_summaries) / total_n success_rate = sum(ws.get("success_rate", 0.0) * ws.get("n", 0) for ws in worker_summaries) / total_n return { "n": total_n, "avg_score": avg_score, "success_rate": success_rate, "workers": len(worker_summaries), "aggregated": True, } def process_alive(pids: list[int]) -> tuple[bool, list[int]]: """Return (any_alive, list_of_alive_pids).""" alive = [] for pid in pids: try: os.kill(pid, 0) alive.append(pid) except (OSError, ProcessLookupError): pass return bool(alive), alive def parse_pids(raw: str) -> list[int]: """Parse a comma-separated list of PIDs.""" return [int(x.strip()) for x in raw.split(",") if x.strip().isdigit()] def load_pids_file(path: Path) -> list[int]: """Load PIDs from a file, one per line or comma-separated.""" if not path.exists(): return [] pids = [] for line in path.read_text().splitlines(): pids.extend(parse_pids(line)) return pids def format_eta(seconds: float) -> str: """Format remaining seconds as a human readable string.""" if seconds < 0 or not seconds: return "unknown" delta = timedelta(seconds=int(seconds)) hours, remainder = divmod(delta.total_seconds(), 3600) minutes, seconds = divmod(remainder, 60) parts = [] if hours: parts.append(f"{int(hours)}h") if minutes: parts.append(f"{int(minutes)}m") if seconds or not parts: parts.append(f"{int(seconds)}s") return "".join(parts) def build_progress_message( *, output_dir: Path, subdir: str, total: int, completed: int, pids: list[int], elapsed_seconds: float, ) -> str: """Build a Feishu progress notification text.""" pct = completed / total * 100 if total else 0.0 avg_per_image = elapsed_seconds / completed if completed else 0.0 remaining = (total - completed) * avg_per_image if completed else 0.0 eta = format_eta(remaining) summary = read_summary(output_dir, subdir) score_line = "" if summary: avg = summary.get("avg_score", 0.0) succ = summary.get("success_rate", 0.0) * 100 score_line = f"\n当前已跑完均分: {avg:.4f} | 成功率: {succ:.1f}%" any_alive, alive_pids = process_alive(pids) pid_line = f"PIDs: {','.join(str(p) for p in pids)}" if pids else "PID: not tracked" status = f"running ({len(alive_pids)}/{len(pids)} workers alive)" if any_alive else "⚠️ all workers gone" return ( f"🤖 Perturb benchmark 进度更新\n" f"目录: {output_dir}/{subdir}\n" f"{pid_line}\n" f"状态: {status}\n" f"进度: {completed}/{total} ({pct:.1f}%)\n" f"已用时间: {format_eta(elapsed_seconds)}\n" f"单图平均: {avg_per_image:.1f}s\n" f"预计剩余: {eta}{score_line}" ) def build_completion_message( *, output_dir: Path, subdir: str, total: int, completed: int, elapsed_seconds: float, aborted: bool = False, ) -> str: """Build completion or abort notification text.""" summary = read_summary(output_dir, subdir) if summary: avg = summary.get("avg_score", 0.0) succ = summary.get("success_rate", 0.0) * 100 score_line = f"均分: {avg:.4f} | 成功率: {succ:.1f}%" else: score_line = "summary.json 尚未生成" if aborted: emoji = "🚨" header = "benchmark 异常退出" extra = f"只完成了 {completed}/{total}" else: emoji = "✅" header = "benchmark 已完成" extra = f"全部 {total}/{total} 完成" return ( f"{emoji} Perturb {header}\n" f"目录: {output_dir}/{subdir}\n" f"{extra}\n" f"总耗时: {format_eta(elapsed_seconds)}\n" f"{score_line}" ) def monitor( *, output_dir: Path, subdir: str, webhook_url: str, interval_seconds: float, pids: list[int], ) -> None: """Main monitoring loop.""" indices_info = read_indices(output_dir) if indices_info is None: # Fallback: wait and retry a few times in case indices.json is not written yet. for _ in range(10): time.sleep(5) indices_info = read_indices(output_dir) if indices_info is not None: break if indices_info is None: msg = f"❌ 监控启动失败:未找到 {output_dir}/indices.json" send_feishu(webhook_url, msg) raise RuntimeError(msg) total, _ = indices_info start_time = time.time() last_report_time = start_time last_completed = 0 # Initial notification send_feishu( webhook_url, f"🚀 Perturb benchmark 监控已启动\n" f"目录: {output_dir}/{subdir}\n" f"总样本数: {total}\n" f"推送间隔: {interval_seconds/60:.0f} 分钟\n" f"开始时间: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}", ) completed = count_completed(output_dir, subdir) while True: time.sleep(min(30, interval_seconds)) now = time.time() elapsed = now - start_time completed = count_completed(output_dir, subdir) # Check if all watched processes died any_alive, alive_pids = process_alive(pids) if pids and not any_alive: aborted = completed < total msg = build_completion_message( output_dir=output_dir, subdir=subdir, total=total, completed=completed, elapsed_seconds=elapsed, aborted=aborted, ) send_feishu(webhook_url, msg) print(f"[monitor] all workers exited. completed={completed}/{total}") return # Periodic progress report if now - last_report_time >= interval_seconds: msg = build_progress_message( output_dir=output_dir, subdir=subdir, total=total, completed=completed, pids=pids, elapsed_seconds=elapsed, ) send_feishu(webhook_url, msg) last_report_time = now # Detect completion if completed >= total: msg = build_completion_message( output_dir=output_dir, subdir=subdir, total=total, completed=completed, elapsed_seconds=elapsed, aborted=False, ) send_feishu(webhook_url, msg) print(f"[monitor] benchmark completed. {completed}/{total}") return # Print local heartbeat if completed != last_completed: print( f"[{datetime.now().strftime('%H:%M:%S')}] completed={completed}/{total} " f"elapsed={format_eta(elapsed)} alive_workers={len(alive_pids)}/{len(pids)}" ) last_completed = completed def main() -> None: parser = argparse.ArgumentParser(description="Feishu monitor for Perturb benchmark") parser.add_argument( "--output-dir", type=Path, default=Path("/workspace/Perturb/benchmark_2000_v15"), help="Benchmark output directory", ) parser.add_argument( "--subdir", type=str, default="v15", help="Subdirectory containing per-image json files", ) parser.add_argument( "--webhook", type=str, default=os.getenv("FEISHU_WEBHOOK_URL", "https://open.feishu.cn/open-apis/bot/v2/hook/b1c64e78-6277-40e6-aed0-795a5c72deb6"), help="Feishu webhook URL (or set FEISHU_WEBHOOK_URL env var)", ) parser.add_argument( "--interval-minutes", type=float, default=30.0, help="Progress report interval in minutes", ) parser.add_argument( "--pid", type=str, default=None, help="Comma-separated process IDs to watch (optional; auto-detected if omitted)", ) args = parser.parse_args() webhook = args.webhook if not webhook: print("[warn] FEISHU_WEBHOOK_URL not set; messages will be printed to stdout only.") pids: list[int] = [] if args.pid: pids = parse_pids(args.pid) else: # Try to auto-detect from pids.txt or legacy pid.txt next to the output dir. candidates = [ args.output_dir.parent / f"{args.output_dir.name}_logs" / "pids.txt", args.output_dir.parent / f"{args.output_dir.name}_logs" / "pid.txt", args.output_dir / "pids.txt", args.output_dir / "pid.txt", ] for cand in candidates: if cand.exists(): pids = load_pids_file(cand) if pids: print(f"[info] auto-detected pids={pids} from {cand}") break monitor( output_dir=args.output_dir, subdir=args.subdir, webhook_url=webhook, interval_seconds=args.interval_minutes * 60, pids=pids, ) if __name__ == "__main__": main()