<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Metrics on YennJ12 Engineering Blog</title><link>https://yennj12.js.org/yennj12_blog_V4/tags/metrics/</link><description>Recent content in Metrics on YennJ12 Engineering Blog</description><generator>Hugo -- gohugo.io</generator><language>en-us</language><lastBuildDate>Tue, 30 Jun 2026 11:00:00 +0800</lastBuildDate><atom:link href="https://yennj12.js.org/yennj12_blog_V4/tags/metrics/feed.xml" rel="self" type="application/rss+xml"/><item><title>Kubernetes Autoscaling Complete Guide (Part 4): Monitoring, Alerting &amp; Threshold Tuning</title><link>https://yennj12.js.org/yennj12_blog_V4/posts/kubernetes-autoscaling-complete-guide-part4-monitoring-alerting/</link><pubDate>Sun, 09 Nov 2025 18:00:00 +0800</pubDate><guid>https://yennj12.js.org/yennj12_blog_V4/posts/kubernetes-autoscaling-complete-guide-part4-monitoring-alerting/</guid><description>Part 4 of the Kubernetes Autoscaling series: Complete guide to monitoring EKS autoscaling with Prometheus and Grafana. Includes CDK setup, alerting rules, custom dashboards, and threshold tuning strategies for production-grade observability.</description></item><item><title>FDE 面試準備指南（十二）：RKK 實戰——AI Agent 統計評估與品質量化</title><link>https://yennj12.js.org/yennj12_blog_V4/posts/fde-interview-guide-part12-agent-evaluation-zh/</link><pubDate>Wed, 03 Jun 2026 11:00:00 +0800</pubDate><guid>https://yennj12.js.org/yennj12_blog_V4/posts/fde-interview-guide-part12-agent-evaluation-zh/</guid><description>你怎麼知道 Agent 可以上線？
直覺不算，「感覺還不錯」不算。
FDE 的工作是把感覺轉成數字，把數字轉成信心——讓客戶的工程團隊能基於證據做決定。
一、核心問題：「夠好」的標準是什麼 Agent 評估的難點不是「怎麼算分」，而是「對誰問什麼問題，要達到什麼分才算夠好」。
三個不同維度的「夠好」：
評估的三個維度（缺一不可） ┌───────────────────┐ ┌───────────────────┐ ┌───────────────────┐ │ 效能（Performance）│ │ 品質（Quality） │ │ 業務（Business） │ │ │ │ │ │ │ │ 快不快？ │ │ 對不對？ │ │ 有沒有用？ │ │ 貴不貴？ │ │ 準不準？ │ │ 用戶滿不滿意？ │ │ │ │ │ │ │ │ tokens/sec │ │ Faithfulness │ │ Task completion │ │ p95 latency │ │ Relevance │ │ User retention │ │ cost/request │ │ Groundedness │ │ Escalation rate │ └───────────────────┘ └───────────────────┘ └───────────────────┘ ↑ ↑ ↑ 系統層關心 工程師關心 客戶關心 只看品質、不看效能：上線後延遲爆炸。</description></item><item><title>ChatPDF RAG 優化（三）：可觀測性與評估 —— Langfuse 追蹤、評估歷史、即時評分</title><link>https://yennj12.js.org/yennj12_blog_V4/posts/chatpdf-rag-optimization-part3-observability-eval-zh/</link><pubDate>Tue, 30 Jun 2026 11:00:00 +0800</pubDate><guid>https://yennj12.js.org/yennj12_blog_V4/posts/chatpdf-rag-optimization-part3-observability-eval-zh/</guid><description>多數 RAG 專案上線後,優化全憑「感覺答案變好了」。 但你說不出 faithfulness 是 0.6 還是 0.9,也不知道上次調 alpha 是讓 nDCG 上升還是下降。 這篇講的就是把「感覺」換成「數字」:追蹤每一次 LLM 呼叫、持久化每一次評估、即時評每一個答案——讓優化變成可被驗證的循環。
一、為什麼可觀測性是 RAG 的最後一哩 第一篇解決切塊與檢索品質,第二篇補上 production 防線。但還缺一塊:你怎麼知道這些優化真的有效?
RAG 的恐怖之處在於它「永遠會給出一個看起來合理的答案」。沒有量測,你根本分不清:
調了 hybrid_alpha,到底是變好還是變壞? 某個答案是基於檢索內容,還是 LLM 自己編的(hallucination)? 這次改動,整體 nDCG@k 的趨勢是上升還是下降? PR #4 補上三層可觀測性:
┌──────────────────────────────────────────────────────┐ │ 三層可觀測性 │ ├──────────────────────────────────────────────────────┤ │ 1. 即時聊天 每個回答附信心分數(faithfulness 等) │ │ 2. 評估工具 每次評估的彙總指標持久化,追蹤趨勢 │ │ 3. Langfuse 所有 LLM 呼叫被追蹤(延遲/token/成本) │ │ ── 設定才開,不設定零開銷 │ └──────────────────────────────────────────────────────┘ 二、Langfuse 追蹤:opt-in 且零開銷 設計原則:不設定 = 完全無感 可觀測性工具最怕的就是「為了觀測而拖慢主流程」。chatPDF 的 Langfuse 整合是完全 opt-in:沒設定環境變數時,它是一個 no-op,零開銷、零風險。</description></item><item><title>Building Centralized Grafana + Prometheus Monitoring with AWS CDK: Multi-Service Observability Platform</title><link>https://yennj12.js.org/yennj12_blog_V4/posts/centralized-grafana-prometheus-monitoring-aws-cdk/</link><pubDate>Sat, 17 Jan 2026 11:00:00 +0800</pubDate><guid>https://yennj12.js.org/yennj12_blog_V4/posts/centralized-grafana-prometheus-monitoring-aws-cdk/</guid><description>Comprehensive guide to architecting a production-ready centralized Prometheus + Grafana monitoring platform using AWS CDK that aggregates metrics from multiple services, clusters, and infrastructure components with federation, remote storage, and advanced alerting.</description></item><item><title>Building a Centralized Monitoring System with AWS CloudWatch and Grafana using CDK</title><link>https://yennj12.js.org/yennj12_blog_V4/posts/centralized-monitoring-system-aws-cloudwatch-grafana-cdk/</link><pubDate>Sun, 30 Nov 2025 14:00:00 +0000</pubDate><guid>https://yennj12.js.org/yennj12_blog_V4/posts/centralized-monitoring-system-aws-cloudwatch-grafana-cdk/</guid><description>🎯 Introduction In distributed systems running on AWS, observability is critical for maintaining reliability, debugging issues, and ensuring optimal performance. A centralized monitoring system provides:
Unified Visibility: Single pane of glass for all services, applications, and infrastructure Proactive Alerting: Detect and respond to issues before they impact users Performance Optimization: Identify bottlenecks and optimization opportunities Cost Management: Track resource utilization and spending patterns Compliance: Meet audit and regulatory requirements for logging Troubleshooting: Quickly diagnose and resolve production issues This comprehensive guide demonstrates how to build a production-ready centralized monitoring system using AWS CloudWatch and Grafana, deployed with CDK (TypeScript).</description></item></channel></rss>