GKRootWire
AI Jensen Huang Says Nvidia Hit AGI, Then Says the Term Is MeaninglessGadgets Sony's New Bravia 6 OLED Targets the Midrange TV FightSecurity Flock Safety Cameras Face Growing Wave of Vandalism as Public Pushback MountsSecurity Report: Nearly 700 AI Agents Coordinated to Compromise Hugging FaceSecurity PaperCut Warns of Zero-Day Flaw Under Active Attack in NG, MF SoftwareSecurity Manchester Airports Group Confirms Hackers Stole Traveler DataAI Jensen Huang Says Nvidia Hit AGI, Then Says the Term Is MeaninglessGadgets Sony's New Bravia 6 OLED Targets the Midrange TV FightSecurity Flock Safety Cameras Face Growing Wave of Vandalism as Public Pushback MountsSecurity Report: Nearly 700 AI Agents Coordinated to Compromise Hugging FaceSecurity PaperCut Warns of Zero-Day Flaw Under Active Attack in NG, MF SoftwareSecurity Manchester Airports Group Confirms Hackers Stole Traveler Data
Cloud & Sysadmin

Cloudflare Trims 100TB of Memory From Its 1.1.1.1 DNS Resolver

A deep dive into cache internals shows how small data-structure tweaks can add up to massive infrastructure savings.

Cloudflare's engineering team recently detailed how they slashed memory usage across the fleet running their 1.1.1.1 public DNS resolver, saving roughly 100 terabytes in aggregate. The gains came not from adding hardware or rearchitecting the service, but from rethinking how cached DNS records are stored in memory at massive scale.

By restructuring the cache's internal data layout, reducing per-entry overhead, and cutting out redundant metadata kept for each cached record, the team squeezed significantly more entries into the same memory footprint. Since 1.1.1.1 handles enormous volumes of queries across thousands of machines worldwide, even tiny per-record savings multiply into huge fleet-wide reductions.

The post walks through the profiling process that surfaced the inefficiencies and the specific engineering trade-offs made to fix them without hurting cache hit rates or lookup speed.

Why it matters: This is a good reminder that at hyperscale, unglamorous data-structure engineering often beats buying more servers. For anyone running caching layers, DNS resolvers, or CDNs, the lesson is that memory overhead per entry compounds brutally once you're serving billions of requests.

Sources: Hacker News