We migrated the Cloudflare Blog to EmDash to prove our stack at massive scale. Here is how we stress-tested performance, safely routed production traffic, and redesigned the frontend experience.
Cloudflare migrated its high-traffic blog to a pre-1.0 CMS running on Workers and absorbed a 28,000 request-per-second DDoS attack on launch week without dropping a single page. 👇
The old blog suffered from unpredictable latency spikes whenever traffic surged from a baseline of 75 requests per second up to viral peaks over 5,000 RPS. Interface quirks were also breaking the reading experience—readers routinely mistook the top-right newsletter sign-up box for a search bar and typed search queries straight into it.
Cloudflare rebuilt the site using EmDash, an Astro-based CMS running natively on Cloudflare Workers, backed by a multi-layer caching strategy and direct database acceleration.
To keep the database from melting under traffic spikes, Cloudflare stacked three distinct caching and connection layers:
Static caching: A Workers Cache layer handles static assets at the edge, serving 99.5% of static files and 70% of total site requests before they ever hit the application logic. Object caching: An EmDash-specific object cache built on Workers KV caches database entities and API responses. Database acceleration: Cloudflare's Hyperdrive maintains pooled, low-latency connections to PlanetScale.
To switch systems safely without downtime, Cloudflare deployed a proxy Worker in front of both platforms. The proxy read a version cookie to direct traffic and automatically fell back to the legacy site if the new system returned a 500 error.
Rather than routing proxy traffic over public DNS, TLS, and external HTTP hops, Cloudflare connected the proxy directly to the new site using internal service bindings (`proxy Worker → new blog Worker`). This allowed them to step traffic up incrementally—from 1% to 5%, 15%, and finally 100%—over a single day.
They also added Model Context Protocol (MCP) endpoints, allowing AI agents to search and fetch posts while letting authors manage content directly through AI tools.
The new architecture flattened the site's P95 latency into a straight line, eliminating the performance spikes of the old system. During its first major event, the blog served 3 million pageviews across 18 posts, handled 850 requests per second under normal load, and weathered the 28,000 RPS attack without visible impact.
NayaRead shows a summary and links to the original article. The full article is hosted on blog.cloudflare.com.