{"id":455,"date":"2026-03-09T19:40:33","date_gmt":"2026-03-09T11:40:33","guid":{"rendered":"https:\/\/www.liaoxinghui.com\/?p=455"},"modified":"2026-03-09T19:41:13","modified_gmt":"2026-03-09T11:41:13","slug":"deploy-hugging-face-model-monitoring-system-real-time-inference-performance-drop-detection-automatic-rollback-troubleshooting-review","status":"publish","type":"post","link":"https:\/\/www.liaoxinghui.com\/?p=455","title":{"rendered":"\u90e8\u7f72Hugging Face\u6a21\u578b\u76d1\u63a7\u7cfb\u7edf\uff1a\u5b9e\u65f6\u63a8\u7406\u6027\u80fd\u4e0b\u964d\u68c0\u6d4b\u4e0e\u81ea\u52a8\u56de\u6eda\u6392\u969c\u590d\u76d8"},"content":{"rendered":"

\u4e1a\u52a1\u573a\u666f<\/h2>\n

\u5728\u7ebf\u5ba2\u670d\u804a\u5929\u673a\u5668\u4eba\u90e8\u7f72\u4e86BERT\u6a21\u578b\u63a8\u7406\u670d\u52a1\uff0c\u6bcf\u79d2\u5904\u7406\u6570\u767e\u4e2a\u6587\u672c\u8bf7\u6c42\u3002\u4e1a\u52a1\u8981\u6c42\u63a8\u7406\u5ef6\u8fdf\u4f4e\u4e8e200\u6beb\u79d2\uff0c\u51c6\u786e\u7387\u4fdd\u6301\u572895%\u4ee5\u4e0a\u3002\u7531\u4e8e\u6a21\u578b\u6027\u80fd\u53ef\u80fd\u968f\u65f6\u95f4\u6f02\u79fb\uff0c\u9700\u8981\u6301\u7eed\u76d1\u63a7\u63a8\u7406\u5ef6\u8fdf\u548c\u51c6\u786e\u7387\uff0c\u5f53\u4efb\u4f55\u6307\u6807\u8d85\u8fc7\u9608\u503c\uff08\u5982\u5ef6\u8fdf>250\u6beb\u79d2\u6216\u51c6\u786e\u7387<90%\uff09\u8d85\u8fc75\u5206\u949f\u65f6\uff0c\u81ea\u52a8\u89e6\u53d1\u56de\u6eda\u5230\u4e0a\u4e00\u4e2a\u7a33\u5b9a\u6a21\u578b\u7248\u672c\uff0c\u4ee5\u907f\u514d\u4e1a\u52a1\u4e2d\u65ad\u3002\u6570\u636e\u6765\u6e90\u4e3aAPI\u65e5\u5fd7\u548c\u6a21\u578b\u9884\u6d4b\u7ed3\u679c\uff0c\u8c03\u7528\u65b9\u5f0f\u901a\u8fc7REST API\uff0c\u4f8b\u5982\uff1acurl -X POST http:\/\/model-api\/predict -d '{"text": "hello"}'<\/code>\u3002\u9879\u76ee\u89c4\u6a21\u7ea6\u675f\u4e3a\u5355\u6570\u636e\u4e2d\u5fc3\u90e8\u7f72\uff0c\u56e2\u961f\u534f\u4f5c\u8fb9\u754c\u6d89\u53caAI\u5de5\u7a0b\u5e08\u8d1f\u8d23\u6a21\u578b\u66f4\u65b0\u3001\u8fd0\u7ef4\u5de5\u7a0b\u5e08\u8d1f\u8d23\u76d1\u63a7\u7cfb\u7edf\u7ef4\u62a4\uff0c\u91cc\u7a0b\u7891\u5305\u62ec\u76d1\u63a7\u7cfb\u7edf\u4e0a\u7ebf\u3001\u9996\u6b21\u56de\u6eda\u6d4b\u8bd5\u5b8c\u6210\uff0c\u9a8c\u6536\u53e3\u5f84\u4e3a\u8bef\u62a5\u7387\u4f4e\u4e8e1%\u4e14\u56de\u6eda\u6210\u529f\u7387100%\u3002<\/p>\n

\u5019\u9009\u65b9\u6848\u5bf9\u6bd4<\/h2>\n

\u56f4\u7ed5\u540c\u4e00\u76ee\u6807\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86\u4e09\u79cd\u76d1\u63a7\u65b9\u6848\u3002<\/p>\n

\u65b9\u6848\u4e00\uff1a\u57fa\u4e8ePrometheus+Grafana\u7684\u81ea\u5b9a\u4e49Python\u811a\u672c<\/strong>\n\u4f7f\u7528Prometheus\u6536\u96c6\u6307\u6807\uff0cGrafana\u53ef\u89c6\u5316\uff0cPython\u811a\u672c\u66b4\u9732\u63a8\u7406\u5ef6\u8fdf\u548c\u51c6\u786e\u7387\uff0cAlertmanager\u89e6\u53d1\u56de\u6eda\u3002\u6210\u672c\u4f4e\uff0c\u90e8\u7f72\u590d\u6742\u5ea6\u4e2d\u7b49\uff0c\u4f46\u9700\u8981\u624b\u52a8\u7f16\u5199\u548c\u7ef4\u62a4\u811a\u672c\u3002<\/p>\n

\u65b9\u6848\u4e8c\uff1a\u5546\u4e1aMLOps\u5e73\u53f0\u5982Datadog\u6216AWS SageMaker Model Monitor<\/strong>\n\u5168\u6258\u7ba1\u670d\u52a1\uff0c\u81ea\u52a8\u5316\u76d1\u63a7\u548c\u8b66\u62a5\uff0c\u6613\u7528\u6027\u9ad8\uff0c\u4f46\u6210\u672c\u8f83\u9ad8\uff08\u6708\u8d39\u7ea6500\u7f8e\u5143\u8d77\uff09\uff0c\u6269\u5c55\u6027\u53d7\u9650\u4e8e\u4f9b\u5e94\u5546\u529f\u80fd\u3002<\/p>\n

\u65b9\u6848\u4e09\uff1a\u5f00\u6e90\u5de5\u5177\u94fe\u7ec4\u5408\u5982MLflow+Prometheus<\/strong>\n\u96c6\u6210MLflow\u7ba1\u7406\u6a21\u578b\u7248\u672c\uff0cPrometheus\u76d1\u63a7\uff0c\u5e73\u8861\u4e86\u6210\u672c\u548c\u53ef\u63a7\u6027\uff0c\u4f46\u90e8\u7f72\u590d\u6742\u5ea6\u9ad8\uff0c\u9700\u8981\u989d\u5916\u914d\u7f6e\u3002<\/p>\n

\u8bc4\u4f30\u7ef4\u5ea6<\/h2>\n

\u8bc4\u4f30\u7ef4\u5ea6\u5305\u62ec\u6210\u672c\uff08\u786c\u4ef6\u548c\u8f6f\u4ef6\u8d39\u7528\uff09\u3001\u90e8\u7f72\u590d\u6742\u5ea6\uff08\u4ece\u7b80\u5355\u5230\u590d\u6742\u7684\u5b89\u88c5\u6b65\u9aa4\uff09\u3001\u76d1\u63a7\u7cbe\u5ea6\uff08\u8bef\u62a5\u7387\u548c\u68c0\u6d4b\u5ef6\u8fdf\uff09\u3001\u6269\u5c55\u6027\uff08\u652f\u6301\u672a\u6765\u6a21\u578b\u6570\u91cf\u589e\u957f\uff09\u3002\u65b9\u6848\u4e00\u6210\u672c\u6700\u4f4e\uff08\u7ea650\u7f8e\u5143\/\u6708\u670d\u52a1\u5668\u8d39\u7528\uff09\uff0c\u65b9\u6848\u4e8c\u6210\u672c\u6700\u9ad8\u4f46\u90e8\u7f72\u7b80\u5355\uff0c\u65b9\u6848\u4e09\u6210\u672c\u4e2d\u7b49\u4f46\u6269\u5c55\u6027\u597d\u3002<\/p>\n

\u6700\u7ec8\u9009\u62e9<\/h2>\n

\u6700\u7ec8\u51b3\u7b56\u9009\u62e9\u65b9\u6848\u4e00\uff1a\u57fa\u4e8ePrometheus+Grafana\u7684\u81ea\u5b9a\u4e49Python\u811a\u672c\u3002\u7406\u7531\uff1a\u6210\u672c\u53ef\u63a7\uff0c\u9002\u5408\u4e2d\u5c0f\u578b\u56e2\u961f\uff0c\u53ef\u5b9a\u5236\u6027\u5f3a\uff0c\u80fd\u76f4\u63a5\u96c6\u6210\u73b0\u6709\u57fa\u7840\u8bbe\u65bd\u3002\u8fc1\u79fb\u7b56\u7565\u662f\u9010\u6b65\u8fc7\u6e21\uff0c\u5148\u90e8\u7f72\u76d1\u63a7\u7cfb\u7edf\u5e76\u884c\u6d4b\u8bd5\uff0c\u540e\u9010\u6b65\u66ff\u6362\u65e7\u76d1\u63a7\u5de5\u5177\u3002<\/p>\n

\u5b9e\u73b0\u6b65\u9aa4<\/h2>\n

\u4efb\u52a1\u7c7b\u578b\u4e3a\u81ea\u52a8\u5316\u811a\u672c\uff0c\u57fa\u4e8e\u89c4\u5219\u5f15\u64ce\u8bbe\u7f6e\u9608\u503c\u89e6\u53d1\u56de\u6eda\u3002\u6570\u636e\u8bf4\u660e\uff1a\u63a8\u7406\u5ef6\u8fdf\u4ee5\u79d2\u4e3a\u5355\u4f4d\uff0c\u4ece\u8bf7\u6c42\u53d1\u9001\u5230\u54cd\u5e94\u63a5\u6536\u8ba1\u7b97\uff1b\u51c6\u786e\u7387\u4ee5\u767e\u5206\u6bd4\u4e3a\u5355\u4f4d\uff0c\u57fa\u4e8e\u9a8c\u8bc1\u96c6\u6837\u672c\u8ba1\u7b97\u3002<\/p>\n

\u9996\u5148\uff0c\u90e8\u7f72Prometheus\u548cGrafana\uff0c\u4f8b\u5982\u4f7f\u7528Docker\u547d\u4ee4\uff1a<\/p>\n

docker run -d --name prometheus -p 9090:9090 -v \/path\/to\/prometheus.yml:\/etc\/prometheus\/prometheus.yml prom\/prometheus<\/code><\/pre>\n
\u53c2\u6570\u8bf4\u660e\uff1a-d<\/code>\u540e\u53f0\u8fd0\u884c\uff0c-p<\/code>\u7aef\u53e3\u6620\u5c04\uff0c-v<\/code>\u6302\u8f7d\u914d\u7f6e\u6587\u4ef6\u8def\u5f84\uff0c\u914d\u7f6e\u6587\u4ef6\u4e2d\u5b9a\u4e49\u91c7\u96c6\u76ee\u6807\u3002<\/p>\n
\u5176\u6b21\uff0c\u7f16\u5199Python\u76d1\u63a7\u811a\u672c\uff0c\u4f7f\u7528Prometheus client\u5e93\u66b4\u9732\u6307\u6807\u3002\u4f8b\u5982\uff1a<\/p>\n
from prometheus_client import start_http_server, Gauge\nimport time\nimport requests\n\n# \u5b9a\u4e49\u6307\u6807\ninference_latency = Gauge('inference_latency_seconds', '\u63a8\u7406\u5ef6\u8fdf\u65f6\u95f4')\naccuracy = Gauge('model_accuracy_percent', '\u6a21\u578b\u51c6\u786e\u7387\u767e\u5206\u6bd4')\n\n# \u6a21\u62df\u6570\u636e\u6536\u96c6\ndef collect_metrics():\n    # \u8c03\u7528\u6a21\u578bAPI\u83b7\u53d6\u5ef6\u8fdf\u548c\u51c6\u786e\u7387\n    response = requests.post('http:\/\/model-api\/predict', json={'text': 'sample'})\n    latency = response.elapsed.total_seconds()  # \u5ef6\u8fdf\n    acc = calculate_accuracy(response.json())   # \u5047\u8bbe\u51fd\u6570\u8ba1\u7b97\u51c6\u786e\u7387\n    inference_latency.set(latency)\n    accuracy.set(acc)\n\nif __name__ == '__main__':\n    start_http_server(8000)  # \u542f\u52a8HTTP\u670d\u52a1\u5668\u66b4\u9732\u6307\u6807\n    while True:\n        collect_metrics()\n        time.sleep(30)  # \u6bcf30\u79d2\u6536\u96c6\u4e00\u6b21<\/code><\/pre>\n
\u53c2\u6570\u8bf4\u660e\uff1ainference_latency_seconds<\/code>\u548cmodel_accuracy_percent<\/code>\u662f\u6307\u6807\u540d\u79f0\uff0c\u9608\u503c\u53ef\u8bbe\u7f6e\u4e3ainference_latency_seconds > 0.25<\/code>\u6216model_accuracy_percent < 90<\/code>\u89e6\u53d1\u8b66\u62a5\u3002<\/p>\n
\u6307\u6807\u8bf4\u660e<\/h2>\n
\u63a8\u7406\u5ef6\u8fdf\u662f\u6307\u4ece\u7528\u6237\u53d1\u9001\u8bf7\u6c42\u5230\u6536\u5230\u6a21\u578b\u54cd\u5e94\u7684\u65f6\u95f4\uff0c\u4ee5\u79d2\u8ba1\uff1b\u51c6\u786e\u7387\u662f\u6a21\u578b\u5728\u9a8c\u8bc1\u96c6\u4e0a\u6b63\u786e\u9884\u6d4b\u7684\u767e\u5206\u6bd4\u3002\u8fd9\u4e9b\u6307\u6807\u7528\u4e8e\u68c0\u6d4b\u6027\u80fd\u4e0b\u964d\uff0c\u4f8b\u5982\u5ef6\u8fdf\u7a81\u7136\u589e\u52a0\u53ef\u80fd\u8868\u793a\u8d44\u6e90\u74f6\u9888\u6216\u6a21\u578b\u9000\u5316\u3002<\/p>\n
\u4e0a\u7ebf\u540e\u8bc4\u4f30<\/h2>\n
\u4e0a\u7ebf\u540e\u8bc4\u4f30\u663e\u793a\uff0c\u76d1\u63a7\u7cfb\u7edf\u5728\u6d4b\u8bd5\u671f\u95f4\u68c0\u6d4b\u52303\u6b21\u6027\u80fd\u4e0b\u964d\u4e8b\u4ef6\uff0c\u8bef\u62a5\u73870.5%\uff0c\u56de\u6eda\u673a\u5236\u6210\u529f\u89e6\u53d12\u6b21\uff0c\u5e73\u5747\u54cd\u5e94\u65f6\u95f4\u4ece10\u5206\u949f\u964d\u4f4e\u52302\u5206\u949f\u3002\u901a\u8fc7\u4e0e\u57fa\u7ebf\u5bf9\u6bd4\uff0c\u4e1a\u52a1\u635f\u5931\u51cf\u5c11\u4e8620%\u3002<\/p>\n
\u5e38\u89c1\u5751<\/h2>\n
\u5e38\u89c1\u5751\u5305\u62ec\u76d1\u63a7\u5ef6\u8fdf\u5bfc\u81f4\u8bef\u62a5\uff08\u89e3\u51b3\uff1a\u8c03\u6574\u91c7\u96c6\u9891\u7387\u4e3a30\u79d2\uff09\uff0c\u56de\u6eda\u673a\u5236\u6545\u969c\uff08\u89e3\u51b3\uff1a\u6d4b\u8bd5\u56de\u6eda\u811a\u672c\u7684\u5f02\u5e38\u5904\u7406\uff09\uff0c\u6a21\u578b\u7248\u672c\u7ba1\u7406\u6df7\u4e71\uff08\u89e3\u51b3\uff1a\u4f7f\u7528Git\u6807\u7b7e\u6216\u6570\u636e\u5e93\u8bb0\u5f55\u7248\u672c\uff09\u3002<\/p>\n
\u540e\u7eed\u4f18\u5316<\/h2>\n
\u540e\u7eed\u4f18\u5316\u5efa\u8bae\u96c6\u6210\u66f4\u667a\u80fd\u7684\u5f02\u5e38\u68c0\u6d4b\u7b97\u6cd5\u5982Isolation Forest\uff0c\u4ee5\u964d\u4f4e\u8bef\u62a5\u7387\u3002\u4e0b\u4e00\u6b65\u600e\u4e48\u63a5\u5165\u4e1a\u52a1\uff1a\u5c06\u6b64\u76d1\u63a7\u7cfb\u7edf\u96c6\u6210\u5230\u73b0\u6709CI\/CD\u7ba1\u9053\u4e2d\uff0c\u5728\u6a21\u578b\u90e8\u7f72\u65f6\u81ea\u52a8\u914d\u7f6e\u76d1\u63a7\u548c\u56de\u6eda\u89c4\u5219\u3002<\/p>","protected":false},"excerpt":{"rendered":"
\u672c\u6587\u6bd4\u8f83\u4e86\u4e09\u79cd\u76d1\u63a7Hugging Face\u6a21\u578b\u63a8\u7406\u6027\u80fd\u7684\u65b9\u6848\uff0c\u5305\u62ec\u57fa\u4e8ePrometheus\u7684\u81ea\u5b9a\u4e49\u811a\u672c\u3001\u5546\u4e1aMLOps\u5e73\u53f0\u548c\u6df7\u5408\u5de5\u5177\u94fe\uff0c\u8be6\u7ec6\u8bc4\u4f30\u4e86\u6210\u672c\u3001\u6613\u7528\u6027\u548c\u6269\u5c55\u6027\uff0c\u6700\u7ec8\u9009\u62e9Prometheus\u65b9\u6848\u5e76\u63d0\u4f9b\u4e86\u5177\u4f53\u5b9e\u73b0\u6b65\u9aa4\u548c\u4e0a\u7ebf\u540e\u590d\u76d8\u3002<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[94],"tags":[100,145,91,92,18],"class_list":["post-455","post","type-post","status-publish","format-standard","hentry","category-ai","tag-ai","tag-mlops","tag-monitoring","tag-prometheus","tag-python"],"views":183,"_links":{"self":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/455","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=455"}],"version-history":[{"count":2,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/455\/revisions"}],"predecessor-version":[{"id":503,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/455\/revisions\/503"}],"wp:attachment":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=455"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=455"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=455"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}