Industries / Technology / Software and SaaS / Cloud and data centers / Web hosting
Content delivery
Keeping copies of a site's files on servers spread around the world, so that pages load quickly, survive sudden crowds and can be shielded from attacks.
6
Milestones achieved, from 1997 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1997
Spreading copies across caches
Six researchers at MIT's Laboratory for Computer Science described "hot spots": sites so swamped with requests that they became unusable, such as NASA's Jet Propulsion Laboratory site after the Shoemaker-Levy 9 comet struck Jupiter, an IBM site during the Deep Blue chess match with Garry Kasparov, and political sites on election night. Their answer, consistent hashing, assigns each page to one of many caches (servers that keep copies) in a way that moves only a small share of pages when a cache is added or removed, and no cache needs a current view of the whole network. They presented it at STOC '97, the ACM theory of computing symposium. Wikipedia lists Akamai's delivery network among the systems that use the method.
Credit
David Karger, Eric Lehman, Tom Leighton, Matthew Levine, Daniel Lewin and Rina Panigrahy (MIT)
2010
Free delivery for any site
On 27 September 2010 Cloudflare opened a public beta at the TechCrunch Disrupt conference. A site joined by changing its domain settings to point at Cloudflare. It kept its existing host and needed no new hardware, software or code. The core service was free; extra features started at $20 a month. Cloudflare said it had run a private beta since June and served more than 1,000 sites. In November 2026 W3Techs, a firm that surveys how websites are built, found Cloudflare in front of 30.0% of all websites, and in front of 86.7% of the sites whose reverse proxy (a service that sits between visitors and a site's own server) it could identify.
Credit
Cloudflare (Matthew Prince, Michelle Zatlyn)
2017
A parser leaks private data
In February 2017 Tavis Ormandy of Google's Project Zero, a team that hunts security flaws, told Cloudflare that some pages passing through its servers came back with stray data attached. A bug in an HTML parser, code that reads and rewrites pages on their way to the visitor, made Cloudflare's edge servers read past the end of a block of memory and send out what lay beyond it: other sites' cookies, login tokens and form contents. Search engines had stored some of it. Cloudflare switched off the first of the three features involved 47 minutes after Google sent details, and worked with search engines to purge the stored copies before it announced the bug on 23 February. By its count the worst period, 13 to 18 February, leaked memory on about 1 in every 3,300,000 requests, and the earliest possible leak was on 22 September 2016. Customers' encryption keys were safe because they were handled by a separate process.
LessonWhen one program rewrites pages for many sites at once, a single bug can hand one site's private data to visitors of another.
Credit
Tavis Ormandy, Google Project Zero (found it); Cloudflare (John Graham-Cumming's report)
2021
One customer's setting, global errors
On 12 May 2021 Fastly, a delivery network, deployed software containing a bug that only a particular customer setting would trigger. At 09:47 UTC on 8 June a customer made a valid change that did, and 85% of Fastly's network began returning errors. Fastly posted on its status page 11 minutes later, found the setting at 10:27 and disabled it; it says 95% of its network was normal within 49 minutes. The BBC reported that gov.uk, the Financial Times, the Guardian, the New York Times, CNN, Twitch and Hulu were among the sites that went down for about an hour.
LessonWhen thousands of sites share one network, a rare input from any one of them can take all of them down.
Credit
Fastly (report by Nick Rockwell); reported by the BBC
2024
A script domain changes hands
Polyfill.js is an open-source library that lets old browsers run newer code. More than 100,000 sites, among them JSTOR and Intuit, loaded it straight from cdn.polyfill.io, by the count of Sansec, a security firm. In February 2024 the domain and the project's GitHub account were sold to a little-known company that Cloudflare names as Funnull. The project's creator, Andrew Betts, said he had never owned the domain and warned users, and Cloudflare set up its own copy of the code that month. On 25 June 2024 Sansec reported that the domain was serving code that sent some mobile visitors to betting sites, only at certain hours and never to site administrators. Within two days Google was blocking ads for shopping sites that used it, Cloudflare was rewriting polyfill.io links on the sites it serves (automatically on its free plan), and the registrar Namecheap had put the domain on hold. Censys, a scanning firm, still counted 384,773 hosts referring to the domain on 2 July.
LessonCode loaded from someone else's domain is only as safe as whoever owns that domain next.
Credit
Sansec (report); Andrew Betts (warning); Cloudflare, Fastly and Namecheap (response)
2025
A doubled file stops Cloudflare
At 11:20 UTC on 18 November 2025 Cloudflare's network began returning error pages for its customers' sites. A permissions change on a database made one query return duplicate rows. That doubled the size of a "feature file" that Cloudflare's bot-detection software reloads every few minutes. The file now held more than 200 features, over a limit in the proxy code, and the software failed on every server that loaded it. The file was rebuilt every five minutes, and only part of the database had been changed, so the network failed and recovered by turns, and engineers at first suspected an attack. Core traffic was largely flowing by 14:30 and errors were back to normal at 17:06. Cloudflare called it its worst outage since 2019, when one badly written regular expression (a text-matching pattern) in a firewall rule used up the processors that serve web traffic on every server for 27 minutes. On 5 December 2025 a second outage hit about 28% of the web traffic Cloudflare serves, for about 25 minutes, by its count. InfoQ, a software news site, reported that both followed configuration changes that reached the whole network almost at once.
LessonA data file sent to every server at once can do as much damage as new code, and needs the same step-by-step rollout.
blog.cloudflare.comblog.cloudflare.comblog.cloudflare.cominfoq.com
Credit
Cloudflare (incident reports); reported by InfoQ
Watch for
Configuration rolled out in stages
Cloudflare says it has finished "Code Orange: Fail Small", the program it started after the November and December 2025 outages. It says that in most cases internal configuration changes no longer reach its network instantly. A new system, Snapstone, releases them in stages while health checks watch for trouble and roll back. Products now fall back to the last good configuration. The Workers runtime is split so that a change reaches a small share of free customers first. It held an outage drill on 7 April 2026 with more than 200 staff and set up a team to write customer updates during incidents. While the program ran, on 20 February 2026, a change to its Bring Your Own IP service withdrew about 1,100 customers' address blocks from the internet; the incident lasted 6 hours and 7 minutes by Cloudflare's count.
Achieved whena monitoring firm outside Cloudflare reports twelve months without a network-wide Cloudflare outage caused by a configuration change, and an incident report shows a bad change stopped at an early stage of its rollout.
Claimed by
Cloudflare, May 2026
Open
01
Purge times measured from outside
A public, repeating test of how fast each large delivery network stops serving an old copy of a file after the site owner asks it to delete its copies (a purge). The builder runs a test site of their own on each network, purges a file on a schedule, and records when each of many places around the world first gets the new copy.
The briefCredit · unclaimed02
Watching domains that serve scripts
A free public feed that ranks the domains from which the most websites load JavaScript, and flags any sign of new control: a new registrant or registrar, new name servers, a certificate from a new issuer, or a script that starts changing for some visitors and not others. A person reviews each flag and labels it.
The briefCredit · unclaimed
Open 01 · the brief
Purge times measured from outside
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, repeating test of how fast each large delivery network stops serving an old copy of a file after the site owner asks it to delete its copies (a purge). The builder runs a test site of their own on each network, purges a file on a schedule, and records when each of many places around the world first gets the new copy.
Why it is still open
The networks publish their own figures. Cloudflare says purges by tag, hostname and prefix take under 150 milliseconds at the median worldwide, and single-file purges 234 milliseconds (September 2024). Fastly's documentation says single-object purges finish across its network "in around 150ms". The only outside measurement found, CDN Planet's Purge Speed Stats, purges one file an hour from three places (the US, the UK and Hong Kong), records whole seconds, and covers CacheFly, CDN77, Cloudflare, Amazon CloudFront and Gcore, not Fastly or Akamai. At whole seconds it cannot test a claim of 150 milliseconds.
Built with what exists
self-serve accounts on Cloudflare, Fastly and Amazon CloudFront, each fronting the builder's own small origin server; the networks' purge APIs; a response header carrying the origin's timestamp, as CDN Planet uses, to tell new copies from old; small virtual machines in 20 or more regions on Fly.io or another cloud, with clocks synchronised by NTP; Python or Go.
First step
Put one test file behind two networks, purge it 100 times while polling from five cloud regions, and compare the result with CDN Planet's figures for a network that both cover.
Done looks like
Three months of hourly purges on at least three networks from at least 20 places, timed to the millisecond, with the raw data and code published and a table setting each network's measured median and 99th percentile beside its own claim.
Open 02 · the brief
Watching domains that serve scripts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A free public feed that ranks the domains from which the most websites load JavaScript, and flags any sign of new control: a new registrant or registrar, new name servers, a certificate from a new issuer, or a script that starts changing for some visitors and not others. A person reviews each flag and labels it.
Why it is still open
The sale of polyfill.io in February 2024 was public, and Cloudflare built a safe copy that month, yet Censys still counted 384,773 hosts referring to the domain a week after it turned malicious. c/side, a company that sells script monitoring, writes that domains serving popular scripts "can dynamically change without notifying anyone or anything", and advises each site to monitor its own pages. Researchers have counted the problem once at a time. In 2012 Nick Nikiforakis and colleagues at KU Leuven and UC Santa Barbara crawled the top 10,000 sites and found 56 script-serving domains, used by 47 of those sites, that anyone could register. In 2023 Takahito Sakamoto and Takuya Murozono of DataSign found that 18 of the 53 domains behind 49 discontinued website tags had been re-registered by others or were free to register. Both were one-off studies. A search for a public watch list of such domains that is kept up to date found none.
Built with what exists
the HTTP Archive's public crawl data in Google BigQuery, which records the scripts that millions of pages load; RDAP, the lookup service for domain registration data; certificate transparency logs through crt.sh; dig for name servers; Python; a small AI model such as Muse Ember to draft a summary of each flag for the reviewer (see the Note "The jobs agents repeat").
First step
Query one month of HTTP Archive data for the 100 domains that serve scripts to the most sites, and record each one's registrar, name servers and certificate issuer as a baseline.
Done looks like
The 1,000 domains that serve scripts to the most sites, checked daily for six months; every change logged in public with its evidence and the reviewer's label; the number of false alarms per week published.