<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[昨日种种，皆成今我。今日种种，皆成新我。]]></title><description><![CDATA[专注 AI Agent 工程化与大模型应用开发，持续分享 Agent 接入、Tool Calling、MCP、RAG、Memory、多 Agent、工作流编排、LangGraph、FastAPI 服务化与生产部署等实战内容。]]></description><link>https://blog.youkuaiyun.com/weixin_71309014</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; weixin_71309014]]></copyright><item><title><![CDATA[近期技术热点：从异常现象到根因定位的完整排查]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170289228</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170289228</guid><author>weixin_71309014</author><pubDate>Sun, 08 Nov 2026 13:30:16 +0800</pubDate><description><![CDATA[导出 ONNX 后推理与 eager 静默不一致，指标下滑无报错。根因是 tracing 按样例输入固化分支，按差异量级分诊后定位控制流。]]></description><category></category></item><item><title><![CDATA[近期技术热点：从异常现象到根因定位的完整排查]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170278545</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170278545</guid><author>weixin_71309014</author><pubDate>Sun, 08 Nov 2026 09:00:15 +0800</pubDate><description><![CDATA[模型导出 ONNX 后输出与 PyTorch 不一致，个别输入会导致预测类别翻转。本文用语义链路查控制流固化、数值链路查浮点差异的双线排查定位根因。]]></description><category></category></item><item><title><![CDATA[近期技术热点：从异常现象到根因定位的完整排查]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170245330</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170245330</guid><author>weixin_71309014</author><pubDate>Sat, 07 Nov 2026 18:00:18 +0800</pubDate><description><![CDATA[文中命令与代码块可直接复制执行，未实测的结论均已单独标注。文中给出了可复现的环境与操作步骤，并标注了哪些结论已经验证、哪些仍属推测。]]></description><category></category></item><item><title><![CDATA[近期技术热点：从异常现象到根因定位的完整排查]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170242152</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170242152</guid><author>weixin_71309014</author><pubDate>Sat, 07 Nov 2026 13:30:17 +0800</pubDate><description><![CDATA[训练恢复后 loss 从 0.31 跳到 0.87，权重与数据排查均无异常。根因是 Adam 矩估计未随权重恢复，等效步长骤增，逐项验证即可定位。]]></description><category></category></item><item><title><![CDATA[近期技术热点：从异常现象到根因定位的完整排查]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170232772</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170232772</guid><author>weixin_71309014</author><pubDate>Sat, 07 Nov 2026 09:00:16 +0800</pubDate><description><![CDATA[模型导出后部分输入推理结果不一致且无报错，排查易误判为数据预处理问题。根因是 `torch.jit.trace` 把数据依赖分支固化为示例输入走过的单一路径。]]></description><category></category></item><item><title><![CDATA[Pod 报 OOMKilled 还是 CrashLoopBackOff？退出码 137 与 1 的排查路径完全不同]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170198080</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170198080</guid><author>weixin_71309014</author><pubDate>Fri, 06 Nov 2026 18:00:23 +0800</pubDate><description><![CDATA[容器反复重启时，137 常被当成 OOM 处理，加内存无效。用 jsonpath 取退出码、用 oom_kill 计数定性，分别指向内核与应用层。]]></description><category></category></item><item><title><![CDATA[DDP多卡eval精度比单卡低3%，先查BN别先怪梯度]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170192572</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170192572</guid><author>weixin_71309014</author><pubDate>Fri, 06 Nov 2026 13:30:18 +0800</pubDate><description><![CDATA[多卡评估精度比单卡低约3个百分点，常被误判为梯度问题。根因是BatchNorm各rank用本地统计量归一化，可同步修复。该修复须在DDP包装前替换BN层，单卡与无进程组场景不宜使用。]]></description><category></category></item><item><title><![CDATA[同一输入两次推理结果不同：PyTorch非确定性三层排查，eval只修第一层]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170184685</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170184685</guid><author>weixin_71309014</author><pubDate>Fri, 06 Nov 2026 09:00:17 +0800</pubDate><description><![CDATA[推理服务同一输入两次返回不同结果，灰度对比与问题复现都失去依据。按现象逐层定位，先确认训练模式残留，再排查算法调优与线程归约差异。严格确定性在部分算子上不可得，需在工程容差与数学一致间取舍。]]></description><category></category></item><item><title><![CDATA[loss 变 nan 无报错：isnan 插桩与 detect_anomaly 谁先谁后]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170146201</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170146201</guid><author>weixin_71309014</author><pubDate>Thu, 05 Nov 2026 18:00:19 +0800</pubDate><description><![CDATA[训练中 `loss` 突然变 `nan` 却无异常堆栈可查，排查此类数值问题最为耗时。先用 `isnan` 统计输入确认 `nan` 首现位置，再用反向异常检测追到出问题的算子。]]></description><category></category></item><item><title><![CDATA[fine-tune后准确率不升反降？LabelEncoder两次fit标签错位才是根因]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170136926</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170136926</guid><author>weixin_71309014</author><pubDate>Thu, 05 Nov 2026 13:30:18 +0800</pubDate><description><![CDATA[fine-tune 后验证准确率反降，常被误判为模型问题，真因是标签编码不一致。排查覆盖从异常到根因的路径，用最小示例复现，并对比统一 fit 与固定类别两种修复。]]></description><category></category></item><item><title><![CDATA[HPA 扩容不生效：为什么设了 10 副本却停在 1 个，先查 metrics-server]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170122292</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170122292</guid><author>weixin_71309014</author><pubDate>Thu, 05 Nov 2026 09:00:21 +0800</pubDate><description><![CDATA[生产集群里 stress-app 跑着 1 个副本，HPA 配了 minReplicas: 1、maxReplicas: 10、CPU 目标 50%。]]></description><category></category></item><item><title><![CDATA[PyTorch训练进Docker静默卡死：strace卡shm调用，默认64MB]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170050613</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170050613</guid><author>weixin_71309014</author><pubDate>Wed, 04 Nov 2026 18:00:20 +0800</pubDate><description><![CDATA[文中命令与代码块可直接复制执行，未实测的结论均已单独标注。文中给出了可复现的环境与操作步骤，并标注了哪些结论已经验证、哪些仍属推测。]]></description><category></category></item><item><title><![CDATA[torch.compile 差异仅 1e-6，边界样本预测为何翻转]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170041633</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170041633</guid><author>weixin_71309014</author><pubDate>Wed, 04 Nov 2026 13:30:18 +0800</pubDate><description><![CDATA[torch.compile 加速后输出与 eager 不一致，部分预测翻转。逐层对比与 graph break 分析定位到动态分支翻转根因。容差回归仅适用于无分支翻转场景，零容忍场景应禁用编译。]]></description><category></category></item><item><title><![CDATA[Docker 容器时间差 8 小时，加了 TZ 环境变量仍不生效的三个根因]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/170022042</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/170022042</guid><author>weixin_71309014</author><pubDate>Wed, 04 Nov 2026 09:00:21 +0800</pubDate><description><![CDATA[容器时间差八小时，补上 TZ 仍输出 UTC，日志全错。三个根因是数据缺失、localtime 未同步、应用自行解析。根治需构建期装 tzdata 并同步 localtime，仅设 TZ 无效。]]></description><category></category></item><item><title><![CDATA[PyTorch显存剩余6GB仍报OOM：降batch无效，根因是碎片化]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/169949042</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/169949042</guid><author>weixin_71309014</author><pubDate>Tue, 03 Nov 2026 18:00:20 +0800</pubDate><description><![CDATA[训练报 CUDA OOM 但显存充足，矛盾常误导排查方向。根因是 caching allocator 碎片化，空闲块够但不连续。适用边界是降 batch 无效且非外进程占用的偶发 OOM。]]></description><category></category></item><item><title><![CDATA[MySQL 加了索引反而更慢：3 万行回表让 Spring Boot 接口从 80ms 涨到 1.4s]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/169939639</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/169939639</guid><author>weixin_71309014</author><pubDate>Tue, 03 Nov 2026 13:30:21 +0800</pubDate><description><![CDATA[接口加索引后从 80ms 涨到 1.4s，并发稍高就拖垮 Pod。用 EXPLAIN ANALYZE 定位 3 万次回表根因，以覆盖索引消除回表。]]></description><category></category></item><item><title><![CDATA[MySQL 8.0 加了联合索引查询反而更慢：3 种失效与回表实测]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/169919768</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/169919768</guid><author>weixin_71309014</author><pubDate>Tue, 03 Nov 2026 09:00:20 +0800</pubDate><description><![CDATA[加了联合索引反而更慢，EXPLAIN 显示 key 命中也难提速。拆解最左前缀断裂、隐式转换与非覆盖回表三类失效并给出可运行示例。排查适用于大表低选择性场景，小表及高选择性索引无需此流程。]]></description><category></category></item><item><title><![CDATA[FastAPI 接口返回 coroutine 对象：漏写 async 时的三种症状与定位]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/169846659</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/169846659</guid><author>weixin_71309014</author><pubDate>Mon, 02 Nov 2026 17:59:22 +0800</pubDate><description><![CDATA[FastAPI 漏写 await 会静默返回 coroutine，副作用丢失。从响应体字符串、日志空白、副作用消失三种症状逐层定位。修复需对齐路由与函数的同步异步声明，行为随版本变化需实测确认。]]></description><category></category></item><item><title><![CDATA[Spring Boot 3.2 偶发500无日志：线程池复用导致请求上下文丢失]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/169838566</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/169838566</guid><author>weixin_71309014</author><pubDate>Mon, 02 Nov 2026 13:53:56 +0800</pubDate><description><![CDATA[线上接口偶发 500，按 traceId 查不到对应异常堆栈。根因是 MDC 未跨线程传递，用 TaskDecorator 在任务前后恢复即可。]]></description><category></category></item><item><title><![CDATA[Alpine 让镜像小了 30%，Java 服务 DNS 解析反而更慢：musl 与 glibc 实测]]></title><link>https://blog.youkuaiyun.com/weixin_71309014/article/details/169835302</link><guid>https://blog.youkuaiyun.com/weixin_71309014/article/details/169835302</guid><author>weixin_71309014</author><pubDate>Mon, 02 Nov 2026 12:32:32 +0800</pubDate><description><![CDATA[Java镜像拖慢CI拉取，换Alpine后DNS解析反而变慢。根因是musl换掉解析器，瘦身应保留glibc并裁剪JRE。无glibc依赖、连接复用、DNS稳定时才值得换Alpine。]]></description><category></category></item></channel></rss>