Hugging Face 自建 CDN:每月传输近 1 EB 应对超大模型下载
Hugging Face 每秒要吐 420 GB 的模型文件,传统 CDN 根本扛不住,他们干脆自己建了一套。文中讲清了 AI 下载和网页下载在缓存设计上的本质区别,还附了一个实时流量地图,做基础设施的工程师会很爱看。
Hugging Face 宣布从今年夏天起自建 CDN,目前每月传输近 1 exabyte,约每秒 420 GB,承载每秒 8 万个请求。官方解释自建原因:传统 CDN 为“大量用户下载大量小文件”设计,而 AI 下载是少量超大文件,且相当一部分请求来自机器而非人。缓存策略因此完全不同:缓存的网页可服务上千访客,而缓存一个 1000 GB 的模型只服务少数客户,但每个客户都要完整、立刻拿到全部内容。
技术上,自建 CDN 的缓存针对超大对象优化,能应对模型发布日的流量洪峰,同时处理多个巨型数据集和热门模型被同时拉取的场景。终端用户也不再是散布在数百个城市的人,而是集中在少数云区域里的 GPU 集群,这改变了“就近缓存”这一核心设计假设。团队还上线了一个实时流量地图页面,可观看全球下载流量的可视化。
对依赖 Hugging Face 拉取模型和数据集的开发者来说,这意味着发布日的大规模下载会更稳定。这一案例也说明 AI 基础设施的流量形态与传统 Web 有本质差异,传统 CDN 厂商需要重新设计面向超大对象的产品。
AI downloads broke the usual CDN model Since this summer, we're running our own CDN at Hugging Face. It's shipping almost 1 exabyte per month, which means, in practice, about 420 GB every second, through 80,000 requests per second. Why did we need to build our own? Well, CDNs were designed to ship many small web files to many people. AI downloads are pretty different: it's a few enormous files, a significant part of them pulled by machines. This changes a lot of design assumptions. You're not caching small objects close to users in hundreds of cities anymore: these users are for a big part GPU clusters sitting in a handful of cloud regions. And a cached web page is useful to 1,000 visitors, while a cached 1,000 GB model is useful to far fewer clients, but each one wants all of it, immediately. So we've built our own CDN, with caches adapted to massive objects, able to handle release-day mega-floods and multiple giant datasets and trending models being pulled at once. The team made a live map of it at huggingface-cdn-live.static.hf.space/index.html Hard to stop watching it - it's quite mesmerizing
本条情报归属 AITOP「产品」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Hugging Face、CDN、开源模型、基础设施,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「Thomas Wolf」(@Thom_Wolf),经 AITOP 多维评分管线结构化提炼。