#!/usr/bin/env bash
# 一键诊断: 收集机器/进程/内存/OOM/日志/数据 全部信息到一个文件, 发回给我分析。
# 用法(在项目根目录):  bash scripts/diag.sh
# 产物:  diag_YYYYmmdd_HHMMSS.txt

cd "$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" || exit 1
OUT="diag_$(date +%Y%m%d_%H%M%S).txt"
PY="$(command -v python || command -v python3)"

# 所有输出同时进文件
exec > >(tee "$OUT") 2>&1

echo "================ 诊断报告 $(date) ================"
echo "主机: $(hostname)  目录: $(pwd)"
echo ""

echo "========== 1. 机器资源 =========="
echo "[内存 free -g]"; free -g
echo ""
echo "[CPU核数] $(nproc)"
echo "[负载 uptime] $(uptime)"
echo ""

echo "========== 2. OOM 检查(夜里有没有被系统杀) =========="
echo "[dmesg killed process]"
( dmesg -T 2>/dev/null || dmesg 2>/dev/null ) | grep -iE "killed process|out of memory|oom-kill" | tail -20 || echo "  (无 / dmesg无权限)"
echo "[/var/log/messages oom]"
grep -iE "killed process|out of memory|oom" /var/log/messages 2>/dev/null | tail -10 || echo "  (无 / 无权限)"
echo ""

echo "========== 3. 进程状态 =========="
echo "[main.py 进程数] $(ps -eo args 2>/dev/null | grep -c '[m]ain.py')"
echo "[ffmpeg 进程数]  $(ps -eo comm 2>/dev/null | grep -c '[f]fmpeg')"
echo "[forkserver/worker] $(ps -eo args 2>/dev/null | grep -c '[m]ultiprocessing\|[f]orkserver')"
echo ""
echo "[吃内存最多的10个进程 RSS]"
ps -eo pid,rss,etime,args --sort=-rss 2>/dev/null | head -11
echo ""

echo "========== 4. Python / 环境 =========="
echo "[python] $($PY --version 2>&1)  路径: $PY"
echo "[which python] $(command -v python 2>/dev/null)"
echo "[自包含python在不在] $(ls python/bin/python3 2>/dev/null && echo yes || echo no)"
echo ""

echo "========== 5. config 关键值 =========="
if [ -f config/config.json ]; then
  $PY - <<'PYEOF'
import json
try:
    d=json.load(open("config/config.json",encoding="utf-8"))
    s=d.get("subtitle",{}); sf=d.get("splitframe",{}); au=d.get("audio",{})
    print("  ocr_workers=%s no_composite=%s merge_workers=%s max_tiles=%s"%(
        s.get("ocr_workers"),s.get("no_composite"),s.get("merge_workers"),s.get("max_tiles_per_composite")))
    print("  cv_detect=%s filter_static=%s filter_watermark_pos=%s split_far_boxes=%s"%(
        s.get("cv_detect"),s.get("filter_static"),s.get("filter_watermark_pos"),s.get("split_far_boxes")))
    print("  min_duration=%s crop_bottom_ratio=%s ffmpeg_threads=%s use_gpu=%s"%(
        s.get("min_duration_seconds"),sf.get("crop_bottom_ratio"),sf.get("ffmpeg_threads"),sf.get("use_gpu")))
    print("  audio.model_size=%s cpu_threads=%s endpoints数=%s"%(
        au.get("model_size"),au.get("cpu_threads"),len(d.get("tencentcloud",{}).get("ocr",{}).get("endpoints",[]))))
except Exception as e:
    print("  读config失败:",e)
PYEOF
else
  echo "  config/config.json 不存在"
fi
echo "[启动命令] $(ps -eo args 2>/dev/null | grep '[m]ain.py' | head -1)"
echo ""

echo "========== 6. app.log 分析 =========="
if [ -f app.log ]; then
  echo "[总行数] $(wc -l < app.log)"
  echo "[关键词计数]"
  for kw in 成功 无字幕 失败 暂停启动 抽帧完成 拼接完成 字幕完成 内存不足 BrokenProcess 崩溃 worker崩 Traceback Error 超时 timeout 冷却 端点; do
    c=$(grep -c "$kw" app.log 2>/dev/null)
    [ "${c:-0}" -gt 0 ] && printf "   %-14s %s\n" "$kw" "$c"
  done
  echo ""
  echo "[最近 40 行]"
  tail -40 app.log
else
  echo "  app.log 不存在"
fi
echo ""

echo "========== 7. 数据.jsonl 分析 =========="
if [ -f 数据.jsonl ]; then
  $PY - <<'PYEOF'
import json,statistics
from collections import Counter,defaultdict
from datetime import datetime,timedelta
latest={}
for line in open("数据.jsonl",encoding="utf-8"):
    line=line.strip()
    if not line: continue
    try: r=json.loads(line)
    except: continue
    v=r.get("video_cos_path")
    if v: latest[v]=r
n=len(latest)
st=Counter(r.get("status") for r in latest.values())
print("  去重视频 %d | 状态 %s"%(n,dict(st)))
if n: print("  成功率 %.1f%%"%(st.get("success",0)/n*100))
# 阶段耗时(成功)
tot=defaultdict(float); te=0
for r in latest.values():
    if r.get("status")!="success": continue
    sd=r.get("stage_durations") or {}; te+=r.get("elapsed_seconds") or 0
    for k in ["split_frame","merge","getsrt"]:
        if k in sd: tot[k]+=sd[k]
if te:
    print("  阶段占比: 抽帧%.0f%% 拼接%.0f%% OCR+字幕%.0f%%"%(tot["split_frame"]/te*100,tot["merge"]/te*100,tot["getsrt"]/te*100))
# OCR/帧
rs=[ic.get("total_ocr_count")/ic.get("frame_count") for r in latest.values() if (ic:=r.get("image_counts") or {}).get("frame_count") and ic.get("total_ocr_count")]
if rs: print("  OCR/帧 中位 %.3f"%statistics.median(rs))
# 时长
durs=sorted((r.get("duration") or 0) for r in latest.values() if r.get("duration"))
if durs: print("  视频时长 中位%.0fs max%.0fs | >20分钟 %d个"%(statistics.median(durs),max(durs),sum(1 for d in durs if d>1200)))
# 单视频平均耗时
els=[r.get("elapsed_seconds") or 0 for r in latest.values() if r.get("status")=="success"]
if els: print("  成功视频平均耗时 %.0fs 中位%.0fs"%(sum(els)/len(els),statistics.median(els)))
# 每小时吞吐 + 有效并发(重叠法)
busy=defaultdict(float); cnt=defaultdict(int)
for r in latest.values():
    try: s=datetime.fromisoformat(r["process_start"]); e=datetime.fromisoformat(r["process_end"])
    except: continue
    if e<=s: continue
    cnt[s.strftime("%m-%d %H")]+=1
    cur=s
    while cur<e:
        hk=cur.strftime("%m-%d %H"); nx=cur.replace(minute=0,second=0,microsecond=0)+timedelta(hours=1)
        seg=min(e,nx); busy[hk]+=(seg-cur).total_seconds(); cur=seg
print("  每小时(完成数 / 有效并发):")
for hk in sorted(busy)[-8:]:
    print("    %s : %d条 / 并发%.1f"%(hk,cnt[hk],busy[hk]/3600))
PYEOF
else
  echo "  数据.jsonl 不存在"
fi
echo ""

echo "========== 8. 磁盘 / GPU =========="
echo "[磁盘 df -h .]"; df -h . 2>/dev/null
echo "[split/ 大小] $(du -sh split 2>/dev/null || echo '无')"
echo "[正在处理的视频数 split/] $(ls split 2>/dev/null | wc -l)"
if command -v nvidia-smi >/dev/null 2>&1; then
  echo "[GPU]"; nvidia-smi --query-gpu=index,memory.used,memory.total,utilization.gpu --format=csv,noheader 2>/dev/null
fi
echo ""
echo "================ 诊断结束 ================"
echo ""
echo ">>> 报告已保存到: $OUT  (把这个文件发回来)"