One customer setting, 85% of a CDN: the Fastly outage of June 2021
A bug shipped on 12 May sat dormant until a valid customer configuration change woke it on 8 June. Fastly says 85% of its network started returning errors.

Somebody changed a setting. That is the whole trigger.
On the morning of 8 June 2021, a Fastly customer made a configuration change to their own service. The change was valid. The platform accepted it, it did what the customer wanted, and there was no reason for anyone to think twice about it.
Within minutes, about 85% of Fastly's network was returning errors, and a long list of the world's most-visited websites stopped loading. The BBC. The New York Times. Reddit. Amazon. GOV.UK, which is where people in the UK go to tax a car or apply for a passport. For roughly an hour, a meaningful slice of the public internet simply was not there.
The cause was not the setting. The cause was a bug that had been sitting in Fastly's software since 12 May, waiting for exactly that setting to come along.
What Fastly says happened
Fastly's published summary is short and, to its credit, does not hedge. A software deployment that began on 12 May 2021 introduced a bug that could be triggered by a specific customer configuration under specific circumstances. The bug was not found during that deployment, nor during the four weeks of normal operation that followed, because nothing exercised the defective path.
Early on 8 June, a customer pushed a valid configuration change that met the conditions. The defect activated, and because edge configuration propagates across a content delivery network quickly by design, it activated almost everywhere at once.
The timeline Fastly published, in UTC:
- 09:47 global disruption begins.
- 09:48 Fastly's monitoring detects it.
- 09:58 a status post goes up.
- 10:27 the triggering customer configuration is identified.
- 10:36 affected services begin recovering.
- 11:00 most services recovered.
- 12:35 incident mitigated.
- 17:25 deployment of the permanent fix begins.
Fastly says that once engineers disabled the triggering configuration, about 95% of the network was operating normally within 49 minutes. One minute from onset to detection is genuinely excellent. Forty minutes from detection to identifying which customer configuration out of many thousands had done it is also fast, and it is the part that would be slow in most organizations.
Error, not attack, and why that was not obvious
This was a software defect. Fastly's account describes a bug and a valid configuration change, with no suggestion of malicious activity anywhere, and nothing in the years since has contradicted that.
It did not look that way at 09:50. A simultaneous, global, near-total failure of a major internet infrastructure provider is exactly what a successful attack on one would look like, and the speculation that morning ran heavily toward attack. The error-versus-attack question is not answerable from symptoms; it is answerable only from evidence inside the failing system. Fastly had that evidence and published a cause within hours, which is the behavior to look for in a provider.
There is a second-order security point worth noting. During the outage, requests were not reaching customer origins at all. For sites that rely on their CDN for web application firewalling, bot management, rate limiting and TLS termination, the security layer and the delivery layer are the same layer. When it is gone, the protection is gone with it. Any origin that is reachable directly, bypassing the edge, is reachable without those controls. That is a reason to make sure your origin only accepts traffic from your provider's network, and to know what happens to your security posture if you ever need to take the edge out of the path in a hurry.
The structural problem: configuration is not code
Most organizations, including sophisticated ones, treat these two things very differently.
Code changes go through review, testing, staging, canary deployment and gradual rollout. There is a well-developed discipline for this, and Google's published material on canarying releases is a reasonable reference for what good looks like.
Configuration changes, by contrast, are frequently designed to take effect immediately and everywhere, because that is the point of them. A CDN exists so that a customer can change caching behavior or routing in seconds rather than waiting for a release cycle. Speed of propagation is a product feature.
The result is a path into production that bypasses every safeguard applied to the code path, and it is exercised constantly by people outside the organization. Fastly's incident is the clearest illustration of where that leads: the defect was in the code, but the thing that reached production without staging was a configuration.
The same pattern recurs across the large platform failures of the last decade. A configuration or policy record, propagating faster than any code change ever would, meeting a code path nobody had tested against it. This is the structural lesson, and it generalizes well beyond content delivery networks. Our cloud security guide makes the related point about guardrails being worth more than care.
Vendor-measured figures versus everything else
It is worth separating what Fastly measured from what was reported around it.
From Fastly: approximately 85% of its network returned errors, approximately 95% of the network operating normally within 49 minutes of disabling the configuration, and the timeline above. These are measurements of Fastly's own edge.
Not from Fastly: the number of end users affected, the number of sites affected, and any financial figure. Fastly has not published those.
From third parties, and worth reading as such: the lists of named websites, which journalists assembled by checking sites and watching outage trackers. Downdetector report counts circulated widely, with around 21,000 user reports for Reddit and over 2,000 for Amazon. Those are counts of people who chose to file a report on a third-party website, which correlate loosely with impact and are not a measurement of it. Press accounts also differ slightly on duration, with several describing roughly an hour of disruption against Fastly's own timeline showing most services recovered by 11:00 and the incident not formally mitigated until 12:35. Both are defensible; they measure different things.
What is confirmed and what is not
Confirmed from Fastly's own summary: the 12 May deployment date, the latent nature of the bug, the valid customer configuration change as the trigger, the 85% figure, the full timeline, and the commitments Fastly made afterwards.
Not published: the identity of the customer, the specific configuration setting involved, the technical nature of the bug itself, or the reason it escaped testing. Fastly said it would publish a deeper post-mortem and review why its quality assurance did not catch the defect, and the level of technical detail available publicly remains the short summary rather than a full engineering analysis. Anyone describing the precise mechanism is going beyond the record.
What to take from it
Stage your configuration changes, even when the platform does not make you. If your provider supports deploying a configuration to a subset of traffic, a single point of presence or a test service first, use it. If it does not, make the change at a time when you are watching, and know how to revert in one action. The ability to revert quickly mattered more here than any amount of pre-change review.
Know what your site does with no CDN in front of it. Not as a theoretical question. Does DNS point at the edge or at the origin? Could you change that, and how long would propagation take? Does the origin accept direct traffic, and if so, what protects it? Would TLS still work? A tested answer is worth having before you need it.
Understand that your edge provider is a security control as well as a delivery mechanism. Consolidating firewalling, bot filtering and rate limiting at the edge is good practice and it concentrates risk. The point is not to unwind it but to know what you lose when it is unavailable, which is the same reasoning that applies across third-party risk generally.
Measure detection and diagnosis separately. Fastly detected the problem in one minute and took forty to identify the cause. Most organizations have the reverse profile: slow to notice, then fast to blame the most recent change. Both numbers are worth knowing about your own operation, and a post-incident review is where you find them.
Have a status channel that does not depend on your own infrastructure. Fastly posted publicly within eleven minutes, which is why customers knew within the hour that this was not their own problem. Being able to tell your users the same thing, from somewhere unaffected, saves a great deal of wasted investigation. It belongs in the preparation section of an incident response plan.
Unexpected "site maintenance" or "verify your account" page during an outage? Opportunists move fast when large sites are unreachable. Paste the link into CyberWatch AI Scan for a free check before entering anything.
The detail that makes this incident instructive is how ordinary both halves were. A routine deployment with a defect in a rarely reached path, which is a normal state for any large codebase. A customer changing a setting they were entitled to change, which happens thousands of times a day. Neither party did anything wrong, and the intersection took out a large fraction of the web for an hour. Resilience work is mostly about narrowing the blast radius of combinations nobody predicted, because the combinations will keep arriving.
Frequently asked questions
What caused the Fastly outage on 8 June 2021?
Fastly says a software deployment that began on 12 May 2021 introduced a bug that could be triggered by a specific customer configuration under specific circumstances. On the morning of 8 June, a customer made a valid configuration change that met those conditions, and the bug activated across the network.
Was it a cyber attack?
No. Fastly's own summary describes an undiscovered software bug triggered by a valid customer configuration change. Nothing in Fastly's account or in subsequent reporting indicates malicious activity.
Was the customer at fault?
No, and Fastly's wording is careful about this. The configuration change was valid. It was permitted by the platform, it did what the customer intended, and it exposed a defect in Fastly's software. A platform that can be broken by a legitimate customer setting has a platform problem.
How much of the network was affected?
Fastly states that about 85% of its network returned errors. That is the company's own measurement of its edge, not an estimate of how many end users were affected, which Fastly has not published.
How long did it last?
Fastly's timeline puts the start of global disruption at 09:47 UTC, detection a minute later, identification of the triggering configuration at 10:27, recovery beginning at 10:36, and most services recovered by 11:00. The incident was mitigated at 12:35 and deployment of the permanent fix began at 17:25.
Which websites went down?
Contemporary reporting named the BBC, the New York Times, the Financial Times, Reddit, Amazon, Spotify, Twitch, Pinterest and the UK government's GOV.UK among affected sites. Those lists were compiled by journalists and outage trackers rather than published by Fastly.
Why does a CDN failure take down a whole site rather than slowing it?
A content delivery network usually sits in front of the origin as the entry point for all traffic, handling TLS termination, routing and caching. If the edge cannot serve a request, the request never reaches the origin, so the site is unreachable even though its servers are healthy.
What is a latent bug?
Code that is defective but never executes the defective path, so it passes testing and runs in production without incident until some input or condition reaches it. Latency of this kind is why a change made weeks earlier can be the cause of today's outage.
What did Fastly say it would do afterwards?
Deploy the permanent bug fix across the network, publish a post-mortem, review why its quality assurance and testing processes did not detect the bug, and work on reducing the time it takes to remediate an event of this kind.
Sources
- Summary of June 8 outage, Fastly
- A vast chunk of the Internet went down because of one company, Fortune
- Major websites begin to come back online after internet outage, Al Jazeera
- Websites around the world experience outages, Rappler
- Canarying Releases, Google SRE Workbook
- Addressing Cascading Failures, Google Site Reliability Engineering
- SP 800-34 Rev. 1, Contingency Planning Guide for Federal Information Systems, US National Institute of Standards and Technology
- Cloud Security Technical Reference Architecture, US Cybersecurity and Infrastructure Security Agency
- Cloud security guidance, UK National Cyber Security Centre


