A browser tab refresh bug kept a server at 100% CPU for 4 days—load scaled with tabs squared
Ok_Negotiation_2587 · reddit · 2026-10-06
The developer of AI Toolbox (a Chrome extension running inside ChatGPT/Claude/Gemini/Grok) recounts a severe production incident: any settings change in one tab triggered every open tab to reload data from the backend, so load grew with the square of tab count—one user with 50 tabs sent 5,751 batches in 28 minutes, keeping the server at 100% CPU for four days.
Why nothing caught it: every piece made sense in isolation, testing with 2–3 tabs behaved fine, and since most code is written by coding agents, neither the agent's review nor the humans' ever asked how many times something runs.
The fix: refresh only the tab that needs it; per-tab cooldown; reload only when data actually changed; a short server-side cache absorbing identical bursts.
The steal-worthy part is a new pre-merge review question for any diff touching network calls: what fires each request, how many times does it run with 1 tab vs 50, can two triggers set each other off—answered with numbers, not 'should be fine'. Asking an agent whether something is efficient gets a yes; asking it to count calls with 50 tabs open surfaces a wrong number or an inability to count—both signals you want.
More from coding & agent
- Laya: Open-Source Model Returns Calibrated Probabilities in 33ms Across 100+ Languages — victormustar · 2026-10-06
- Proactivity-Gym: 23-agent study shows correct work alone can't sustain user trust — minnesotanlp · 2026-10-06
- OpenRUA: off-the-shelf Claude Code drives robots via ROS 2 at 99% success, zero training — Zhaoyang Chu · 2026-10-06
- SearchJev makes search-agent decisions 5x faster with 41-74% lower calibration error — Congfeng Cao · 2026-10-06
- LangChain pitches Managed Deep Agents as a way to monetize personal agent workflows — hwchase17 · 2026-10-06
- Dev envisions local agent 'rooms' no company owns and no server can read — RileyRalmuto · 2026-10-06