首次提交

tools/查端点存活.py

@@ -0,0 +1,147 @@
+# -*- coding: utf-8 -*-
+"""
+逐个探测 config 里的 OCR 端点, 看有多少台【真的在服务】。
+用来确认"120台只活了20台"这类问题——死端点越多, 活端点被压得越狠 → 超时。
+
+用法:
+  python3 tools/查端点存活.py                 # 自动找一张真实帧图做OCR探测
+  python3 tools/查端点存活.py --tcp            # 只测TCP端口连通(更快, 不发OCR)
+  python3 tools/查端点存活.py --timeout 5
+  python3 tools/查端点存活.py --list-dead       # 额外列出死端点IP
+"""
+import argparse
+import base64
+import glob
+import os
+import socket
+import sys
+import time
+from collections import Counter
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+import requests
+from config.config_utils import config
+
+OCR_PATH = "/youtu/ocrapi/hpgeneralocr_v3"
+DEFAULT_PORT = 18081
+
+
+def norm(e):
+    e = str(e).strip().rstrip("/")
+    if not e:
+        return None
+    if not e.startswith("http"):
+        e = "http://" + e
+    host = e.split("://", 1)[1]
+    if ":" not in host:
+        e = e + ":" + str(DEFAULT_PORT)
+    return e
+
+
+def host_port(ep):
+    h = ep.split("://", 1)[1]
+    host, _, port = h.partition(":")
+    return host, int(port or DEFAULT_PORT)
+
+
+def find_image_b64():
+    """找一张真实帧图转 base64; 找不到返回 None(改用TCP探测)。"""
+    for pat in ("./split/**/*.jpg", "./composite/**/*.jpg", "./split/*/*.jpg"):
+        for p in glob.glob(pat, recursive=True):
+            try:
+                with open(p, "rb") as f:
+                    return base64.b64encode(f.read()).decode()
+            except OSError:
+                continue
+    return None
+
+
+def probe_ocr(ep, img_b64, timeout):
+    url = ep + OCR_PATH
+    payload = {"app_id": str(int(time.time() * 1000)), "image": img_b64,
+               "ocr_template": "OCR", "options": {"pdf_page_index": 0, "detect_text": True}}
+    t0 = time.time()
+    try:
+        r = requests.post(url, json=payload, timeout=timeout,
+                          headers={"Content-Type": "application/json"})
+        dt = time.time() - t0
+        return (ep, "alive", dt, "") if r.status_code == 200 else (ep, "http_err", dt, f"HTTP {r.status_code}")
+    except requests.exceptions.ConnectTimeout:
+        return (ep, "dead", time.time() - t0, "连接超时")
+    except requests.exceptions.ReadTimeout:
+        return (ep, "slow", time.time() - t0, "读超时(连得上不回)")
+    except requests.exceptions.ConnectionError:
+        return (ep, "dead", time.time() - t0, "拒绝/不可达")
+    except Exception as e:
+        return (ep, "dead", time.time() - t0, str(e)[:40])
+
+
+def probe_tcp(ep, timeout):
+    host, port = host_port(ep)
+    t0 = time.time()
+    try:
+        with socket.create_connection((host, port), timeout=timeout):
+            return (ep, "alive", time.time() - t0, "TCP通")
+    except socket.timeout:
+        return (ep, "dead", time.time() - t0, "TCP超时")
+    except Exception as e:
+        return (ep, "dead", time.time() - t0, str(e)[:40])
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--timeout", type=float, default=5.0)
+    ap.add_argument("--tcp", action="store_true", help="只测TCP端口连通(不发OCR)")
+    ap.add_argument("--list-dead", action="store_true")
+    args = ap.parse_args()
+
+    eps = config.get_tencent_ocr_config().get("endpoints") or []
+    eps = [u for u in (norm(e) for e in eps) if u]
+
+    img = None
+    if not args.tcp:
+        img = find_image_b64()
+        if img is None:
+            print("没找到帧图, 改用 TCP 连通探测(--tcp)。")
+    mode = "TCP连通" if (args.tcp or img is None) else "真实OCR请求"
+    print(f"探测 {len(eps)} 个端点 (方式: {mode}, 超时 {args.timeout}s)...")
+
+    def work(e):
+        if args.tcp or img is None:
+            return probe_tcp(e, args.timeout)
+        return probe_ocr(e, img, args.timeout)
+
+    res = []
+    with ThreadPoolExecutor(max_workers=64) as ex:
+        for f in as_completed([ex.submit(work, e) for e in eps]):
+            res.append(f.result())
+
+    c = Counter(r[1] for r in res)
+    alive = [r for r in res if r[1] in ("alive", "http_err")]
+    lats = sorted(r[2] for r in alive)
+    total = len(eps)
+    good = c["alive"] + c["http_err"]
+    print("=" * 52)
+    print(f"  可用(alive/http_err): {good}")
+    print(f"  慢(读超时):           {c['slow']}")
+    print(f"  死(连不上/拒绝):      {c['dead']}")
+    if lats:
+        print(f"  可用端点延迟: 中位 {lats[len(lats)//2]*1000:.0f}ms  最慢 {lats[-1]*1000:.0f}ms")
+    print("=" * 52)
+    print(f"结论: {total} 个端点里 {good} 个可用 ({good/total*100:.0f}%)")
+    if good < total * 0.8:
+        print(f"⚠️ {total-good} 个端点不可用! 负载全压在 {good} 个上 → 这极可能就是OCR超时根因。")
+        print(f"   把死端点服务拉起 / 或从config删掉死的, OCR容量最高可提升 {total/max(good,1):.1f} 倍。")
+    else:
+        print("端点基本都活着 → OCR超时不是'端点少', 是'整体容量/CPU抢占'问题。")
+
+    if args.list_dead:
+        print("\n死/慢端点:")
+        for ep, st, dt, msg in sorted(res, key=lambda x: x[1]):
+            if st in ("dead", "slow"):
+                print(f"  {st:6s} {ep:32s} {msg}")
+
+
+if __name__ == "__main__":
+    main()