惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

腾讯CDC
博客园 - Franky
MyScale Blog
MyScale Blog
L
LangChain Blog
Martin Fowler
Martin Fowler
Recent Announcements
Recent Announcements
Stack Overflow Blog
Stack Overflow Blog
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
博客园 - 司徒正美
量子位
A
About on SuperTechFans
C
Check Point Blog
大猫的无限游戏
大猫的无限游戏
Last Week in AI
Last Week in AI
小众软件
小众软件
Apple Machine Learning Research
Apple Machine Learning Research
I
InfoQ
V
Visual Studio Blog
Vercel News
Vercel News
B
Blog
爱范儿
爱范儿
aimingoo的专栏
aimingoo的专栏
U
Unit 42

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
A one-line cache key bug cost me $187/month and leaked ad...
강해수 · 2026-06-26 · via DEV Community

강해수

60% of my $312 Anthropic bill last month came from a single bug: an MCP router cache key that was missing a tenant ID.

The fix was literally this:

// before
const cacheKey = `mcp:context:${requestId}`;

// after
const cacheKey = `mcp:context:${tenantId}:${requestId}`;

That one missing segment meant warm Cloudflare Worker instances were serving cached Vectorize results from advertiser A into advertiser B's tool responses. In a production ad analytics SaaS. Not a demo.

The counterintuitive part: I assumed V8 isolate boundaries protected me. They don't — not in the way most people think. Isolate-level isolation applies between separate Worker deployments, not between two concurrent requests hitting the same warm Worker instance. Module-scope variables survive across requests. So any context manager or cache object you initialize at module level is shared state, even on Workers.

The failure mode was subtle enough to take 6 weeks to find. Vectorize query volume was 3× expected — that was the first signal. Digging into logs, I found cache hits for tenant a9f2 being served to sessions belonging to tenant b3c1. The corrupted cache contained vector search results, so every bad hit triggered a downstream re-fetch chain. That cascade is what blew up the token spend: wrong cache data → Claude retries with fresh context → Sonnet input tokens accumulate fast.

After fixing the cache key namespace and adding a PostToolUse hook that throws on tenant ID mismatch in tool response metadata, Sonnet input costs dropped from ~$187/month to ~$94. Vectorize queries fell ~40% over the same period.

One thing worth flagging for anyone on a similar stack: this specific fix — scoping everything to Workers' ExecutionContext per request — doesn't translate cleanly to long-running Node processes on something like Fly.io. There, AsyncLocalStorage is the right primitive. Porting the Workers pattern directly will give you a false sense of safety.

I wrote up the full breakdown — including the PostToolUse hook implementation, the KV/D1 cache key enforcement pattern, and the cases where this isolation design is overkill — over on riversealab.com.

Full post →