On November 13, 2026, excess network traffic disrupted the FAA’s air traffic control network and stopped departures across the United States from 3:36 PM ET, when the FAA announced the stop according to CBS, until about 8:15 PM ET. According to the FAA, the traffic was traced to two sites, a radio site in El Dorado, Arkansas with a newly installed router and a radar site in Amarillo, Texas, and the root cause is still under investigation. The lesson for any engineering team is that redundancy helps only when the backup does not share the primary’s failure causes, so failure domains have to be small and separated.
Disclosure: Control Plane publishes this guide.
TL;DR
According to the FAA and the U.S. Department of Transportation, excess traffic traced to two sites destabilized the air traffic control network and led to a nationwide ground stop of about 4 hours 40 minutes. The root cause is still under investigation.
- According to the FAA, about 1,000 connection attempts every four to five seconds came from a radio site in El Dorado, Arkansas and a radar site in Amarillo, Texas, causing instability at 17 of the FAA’s 22 Air Route Traffic Control Centers that spread into several approach control facilities.
- It was not a full outage. The FAA said some frequencies and radar feeds stayed operational, and life-saving, law enforcement, and military flights continued.
- Homeland Security said there were no indications of a cyberattack. Transportation Secretary Sean Duffy said “we don’t believe this was cyber at all.”
- The instability spread across the network from the two source sites. Officials have not published the network design or how the traffic reached other sites.
- When a backup shares a path with the primary, both can fail together, a pattern engineers call correlated failure.
- The software equivalents are retry storms, connection floods, resource exhaustion, and correlated failures in systems that look redundant but share a dependency.
- Check whether one component can reach every other, whether primary and backup share a network, whether retries are rate-limited, and whether you can isolate a source within minutes.
- Control Plane runs each workload in every location of Control Plane’s Global Virtual Cloud (GVC), and every location serves traffic, so when a workload runs in two or more locations and one region fails, users are routed to the healthy locations.
What caused the FAA ground stop on November 13, 2026?
The FAA and the Department of Transportation have published the sequence of events, but they have not released exact times for most steps.
| Time (ET) | Event |
|---|---|
| Not published | L3Harris performs a copper-to-fiber replacement at a radio site in El Dorado, Arkansas, where a new router is installed |
| Not published | The network is unstable for about 45 minutes while engineers cannot identify the cause |
| 3:36 PM | The FAA announces a nationwide ground stop (CBS); roughly 3,500 flights are already airborne |
| Not published | L3Harris blocks the two source locations, and the network stabilizes |
| About 8:15 PM | Departures resume, about 4 hours 40 minutes after the stop |
Several questions remain open. According to the FAA, the root cause is still under investigation by the FAA and L3Harris. On November 17, Duffy said “the root cause of what happened, we don’t have clarity yet,” according to the AP. HPE, which owns Juniper, says it believes the routers were working properly. The FBI is examining the router, and the White House formed a tiger team. The FAA disabled the same function on more than 60 other Juniper routers and paused installing them.
According to NBC, it was the second nationwide ground stop since September 11, 2001. The previous one was in January 2023, when the FAA’s pilot-notice (NOTAM) system failed. For context on how often large infrastructure outages happen, the cloud outage tracker logs major provider incidents with sourced durations.
How did a network flood stop flights nationwide?
The table maps the published facts, and the questions officials have not yet answered, to their counterparts in distributed software and the controls that address them.
| Stage | What happened at the FAA | The software equivalent | The control that stops it |
|---|---|---|---|
| Work at a source site | According to the FAA, L3Harris was doing a copper-to-fiber replacement at El Dorado, where a new router was installed; officials have not said whether this work played a role | A deploy or configuration change pushed to production | Roll the change to one cell first, with automated rollback thresholds |
| Repeated connection attempts flood neighbors | The FAA traced excess connection attempts to the El Dorado and Amarillo sites | A retry storm or connection flood from one service | Rate limits with exponential backoff and jitter on outbound connections |
| Neighbors absorb the traffic | Officials have not published how the other routers were affected internally | Resource exhaustion in services that share a thread pool or connection pool | Per-client quotas, connection limits, and backpressure |
| The fault reaches many sites | According to the FAA, instability reached 17 of 22 Air Route Traffic Control Centers and spread into several approach control facilities | A fault that crosses service or region boundaries | Segmentation, bulkheads, and blast-radius limits set at design time |
| Whether primary and backup share a path | The FAA said some frequencies and radar feeds stayed operational; it has not published whether backup paths carried the same traffic. | Primary and standby in one availability zone, or both behind one load balancer | Primary and backup in separate failure domains with separate networks |
| Isolating the source | According to the FAA, the network stabilized after L3Harris blocked the two source locations | Time to detect and cut off a misbehaving service | Alerts on abnormal traffic trigger automated or rehearsed blocks that cut off the source |
The pattern holds for physical networks and for software. One component misbehaves. Its neighbors absorb the load. If nothing bounds the traffic, the problem can spread. If the backup shares the same dependency, it can fail for the same reason as the primary.
Why can redundancy fail to prevent an outage like this?
Officials have not published the FAA network design. The general pattern is well understood.
Engineers call this correlated failure. Two components are correlated when one cause can take down both. A primary path and a backup path that run through the same routers will both feel a flood on those routers. A primary database and a standby replica in one availability zone will both go offline if that zone loses power. In each case the backup exists, but it sits inside the same failure domain as the primary.
A failure domain is the set of components one fault can reach. A large network that connects every site to every other site has one large failure domain unless something divides it. Segmentation, per-source rate limits, and separate paths for backup traffic are the usual dividers. Each adds cost and operational work.
Shared dependencies tend to accumulate over years. They can sit behind labels like “redundant” and “highly available” for a long time. They usually show up only when a fault exercises the exact path they share.
Redundancy and resilience answer different questions. Redundancy asks whether a backup exists. Resilience asks whether that backup survives the same fault that takes down the primary. Officials have not published which parts of the FAA network shared exposure to the traffic or how it reached 17 facilities, and they have not said whether the findings will be published in this detail.
The guide on designing systems that survive the inevitable covers architectural patterns for keeping backups out of the primary’s failure domain.
How do you keep a failure in its box?
Engineering teams can use these eight questions in design reviews and readiness checks.
| Question | Why it matters | Good answer |
|---|---|---|
| Can one component reach every other? | A single misbehaving component can then flood or destabilize the whole system | Components sit in cells or zones with explicit, limited connectivity between them |
| Do primary and backup share a network, power, or management plane? | If the backup shares that dependency, it can go down with the primary | Primary and backup use separate paths, carriers, and management networks |
| Are retries rate-limited with backoff? | Unlimited retries turn a short error into a sustained flood | Retry logic uses exponential backoff, jitter, and a retry cap enforced at the client |
| Can you isolate a source within minutes? | A slow response lets the fault reach more components | Alerts on abnormal traffic trigger automated or rehearsed blocks that cut off the source |
| Are changes rolled out to one cell first? | A change applied everywhere at once has a blast radius equal to the whole system | Changes go to one cell, soak for a set period, and promote only when health metrics hold |
| Is there a per-source limit on new connections? | A flood of connection attempts is a common failure mode in networks and in software | Each source has a connection rate limit, and traffic above it is rejected automatically |
| Does your status and alerting path survive the failure? | If monitoring rides the failing network, you lose visibility when you need it most | Monitoring uses a separate path, and alerts reach on-call engineers during a full production outage |
| Have you tested losing a whole region? | Independence between regions is an assumption until a test removes one | The team runs regular game days that remove a full region and measures recovery time |
What does this mean for cloud architecture?
Cloud providers split infrastructure into regions and availability zones so that each one is a separate failure domain. A region is a geographic area with several data centers. An availability zone is one or more separate data centers inside a region, with its own power, cooling, and networking. A fault in one zone is designed to stay in that zone.
That design helps only when your application keeps the separation. An application spread across three zones can still depend on one load balancer, one DNS provider, or one management API. Those shared services put all three zones in one failure domain.
Cell-based architecture applies the same idea inside the application. You divide the system into self-contained cells, and each cell serves a subset of users or traffic. A fault in one cell stays there while other cells keep serving. This is the bulkhead pattern, named for the walls that keep one flooded compartment from sinking a ship.
Rate limits cap how much traffic a caller sends, and circuit breakers stop calls to a dependency that is failing. Every service-to-service call should cap connection rate, retry rate, and outstanding requests. When a downstream service slows or returns errors, the circuit breaker opens and stops sending traffic. That keeps the caller from exhausting its own resources and passing the failure upstream.
Active-active deployment across separate regions or providers protects against correlated regional failure. Workloads run in several regions at once, and each region serves traffic on its own. When one region fails, traffic shifts to the regions that are still healthy. The separation holds only if the regions do not share a global database, a single DNS authority, or one management API that all of them need to serve requests.
The guide on surviving a cloud provider outage without a runbook covers the operational side of losing a region.
How does Control Plane fit?
Control Plane handles Day 2 operations for you, including autoscaling, built-in observability, and security. A Global Virtual Cloud (GVC) is Control Plane’s named set of locations across AWS, GCP, Azure, and Oracle Cloud Infrastructure regions, plus Kubernetes clusters you already run and your own servers. A workload runs in every location of its GVC, and every location serves traffic.
When a workload runs in two or more locations and one region fails, users are routed to the healthy locations. On October 20, 2025, 20% of Control Plane customers had AWS us-east-1 in their GVC, and their workloads were “active and running in healthy locations within 10 seconds,” with no customer downtime, according to Control Plane’s internal incident telemetry.
Control Plane runs the workloads you deploy on it; it does not change how third-party SaaS, identity providers, or other organizations’ networks behave.
FAQ
These answers cover common questions about the November 13, 2026 FAA ground stop and the failure-domain lessons it raises.
Why did the FAA stop all flights on November 13, 2026?
According to the FAA, excess network traffic of about 1,000 connection attempts every four to five seconds disrupted the air traffic control network. The FAA traced the traffic to two sites, a radio site in El Dorado, Arkansas where a new router had been installed and a radar site in Amarillo, Texas. The root cause is still under investigation. The stop ran from 3:36 PM ET, when the FAA announced it according to CBS, to about 8:15 PM ET, and roughly 3,500 flights were already airborne when it began.
Was the FAA ground stop a cyberattack?
Officials have found no sign of one. Homeland Security said there were no indications of a cyberattack, and Transportation Secretary Sean Duffy said “we don’t believe this was cyber at all.” The FBI is examining the router. The root cause is still under investigation, and HPE, which owns Juniper, says it believes the routers were working properly.
How can one source of excess traffic disrupt a large network?
One source can disrupt a large network when it floods its neighbors with connection attempts. Per-source rate limits and segmentation are the usual controls that keep that traffic local. Officials have not published the internal details of how the traffic reached 17 facilities.
What is a failure domain?
A failure domain is the set of components that a single fault can affect. Well-designed systems keep failure domains small and separate, so a problem in one area stays there. In cloud architecture, availability zones and regions are built to act as separate failure domains.
What is a correlated failure?
A correlated failure happens when components meant to be independent fail from the same cause. They usually share a hidden dependency, such as a network path, a power source, a management network, or a software version. A primary and backup that share such a dependency give less protection than the label “redundant” suggests.
What is cell-based architecture?
Cell-based architecture splits a system into self-contained units called cells, each serving a subset of users or traffic. A fault in one cell stays in that cell while the others keep serving. It is the software form of the bulkheads that divide a ship’s hull.
How do companies prevent cascading failures?
Companies prevent cascading failures with a mix of architecture and operations. On the architecture side they use segmentation, cells, separate failure domains for primary and backup, per-source rate limits, and circuit breakers. On the operations side they use staged rollouts, alerts on abnormal traffic, rehearsed isolation steps, and game days that remove a full region.
How many nationwide ground stops have there been since 9/11?
According to NBC, the November 13, 2026 ground stop was the second nationwide ground stop since September 11, 2001. The previous one was in January 2023, when the FAA’s pilot-notice (NOTAM) system failed.
What is the FAA doing to prevent future ground stops?
L3Harris blocked the two source locations, El Dorado and Amarillo, and the network stabilized. The FAA disabled the same function on more than 60 other Juniper routers and paused installing them. The White House formed a tiger team, and the FBI is examining the router. Duffy is asking for about $50 billion for air traffic control modernization, according to the AP.
How does running in multiple regions reduce the risk of cascading failures?
Running a workload in several regions at once means one regional failure leaves other regions serving users. Control Plane runs a workload in every location of its GVC across AWS, GCP, Azure, and Oracle Cloud Infrastructure regions. When the workload runs in two or more locations and one region fails, users are routed to the healthy locations.
Talk to the Control Plane team about running workloads in separate regions with traffic already flowing, or sign up free to start.
Originally published on the Control Plane blog.
