Summary

On August 21, 2025, excessive traffic aimed at clients hosted in AWS us-east-1 saturated the links between Cloudflare and that region. Connections to and from Cloudflare via AWS us-east-1 degraded into high latency, packet loss and origin failures. Customers with origins in us-east-1 first saw impact at 16:27 UTC; it was substantially reduced by 19:38 UTC, with intermittent latency spikes persisting until 20:18 UTC.

The event was regional: global Cloudflare services were unaffected and degradation stayed confined to traffic between Cloudflare and AWS us-east-1. The cause was a traffic surge from a single customer that overloaded the links — a congestion event, not an attack and not a BGP hijack.

How Cloudflare reaches origins

Most customers place Cloudflare in front of origin servers it does not operate, acting as a reverse proxy. Requests are inspected first: cached content or content configured for direct serving is delivered without contacting the origin. Otherwise Cloudflare fetches from the origin, serves it, and caches it along the way when eligible, so later requests avoid the round trip.

Cached responses travel over internal Data Center Interconnect (DCI) links through a series of network equipment until they reach the edge routers.

BLOG-2938 2

Internal capacity is provisioned above the available traffic demand at a location to absorb redundant-link failures, failover from other locations, traffic engineering, or user surges. Most Cloudflare links were behaving normally; the problem was insufficient capacity on some edge router links to an AWS peering switch for this particular surge.

Sequence of events

At roughly 16:27 UTC, a customer began sending many requests from AWS us-east-1 to Cloudflare for cached objects. The response traffic saturated every available direct peering connection between Cloudflare and AWS. AWS then withdrew some BGP advertisements to Cloudflare over congested links in an attempt to relieve the pressure; that rerouted traffic onto additional peering links reachable through an offsite network interconnection switch, which also saturated, causing significant degradation.

Two conditions deepened the impact: one direct peering link was running at half-capacity because of a pre-existing failure, and the DCI connecting Cloudflare's edge routers to the offsite switch was due for a capacity upgrade.

BLOG-2938 3

The incident team engaged AWS immediately, and joint work relieved the congestion and fully restored service for affected customers.

Timeline and traffic behavior

Time

Description

2025-08-21 16:27 UTC

Traffic surge for single customer begins, doubling total traffic from Cloudflare to AWSIMPACT START

2025-08-21 16:37 UTC

AWS begins withdrawing prefixes from Cloudflare on congested PNI (Private Network Interconnect) BGP sessions

2025-08-21 16:44 UTC

Network team is alerted to internal congestion in Ashburn (IAD)

2025-08-21 16:45 UTC

Network team is evaluating options for response, but AWS prefixes are unavailable on paths that are not congested due to their withdrawals

2025-08-21 17:22 UTC

AWS BGP prefixes withdrawals result in a higher amount of dropped trafficIMPACT INCREASE

2025-08-21 17:45 UTC

Incident is raised for customer impact in Ashburn (IAD)

2025-08-21 19:05 UTC

Rate limiting of single customer causing traffic surge decreases congestion

2025-08-21 19:27 UTC

Network team additional traffic engineering actions fully resolve congestionIMPACT DECREASE

2025-08-21 19:45 UTC

AWS begins reverting BGP withdrawals as requested by Cloudflare

2025-08-21 20:07 UTC

AWS finishes normalizing BGP prefix announcements to Cloudflare over IAD PNIs

2025-08-21 20:18 UTC

IMPACT END

When impact began, traffic tied to one customer stood out and produced congestion.

BLOG-2938 4

Both parties applied manual traffic actions. AWS's mitigation attempts are visible in the count of IP prefixes it advertised to Cloudflare during the outage — the dips show prefixes withdrawn from BGP sessions to steer traffic elsewhere.

BLOG-2938 5

Queues on the routers grew substantially and began dropping packets. Edge routers dropped high-priority packets consistently throughout the outage.

BLOG-2938 6

Customers primarily experienced latency, loss (timeouts) or low throughput. Latency Service Level Objectives that mimic customer requests back to origins and measure availability and latency fell below the acceptable threshold in lockstep with the packet drops.

BLOG-2938 7

After congestion eased, AWS and Cloudflare both worked to normalize the prefix advertisements they had adjusted during mitigation. That produced a long tail of latency for some customers, explaining why packet drops clear before customer latency recovers.

Remediations

The incident underlined the need for safeguards so one customer's usage cannot degrade the wider ecosystem. The takeaways: architect for better customer isolation so no single entity can monopolize shared resources, and expand network capacity to meet demand.

Short- and medium-term actions:

  • Build a mechanism to selectively deprioritize a customer's traffic when it congests the network enough to affect others.
  • Expedite the DCI upgrades to provide capacity significantly above today's levels.
  • Work with AWS so their BGP traffic engineering actions and Cloudflare's do not conflict.

Longer term, Cloudflare plans an enhanced traffic management system that allocates network resources per customer, creating a budget that, once exceeded, prevents that customer's traffic from degrading service for anyone else. The system will also automate many of the manual actions taken during this incident.

Outcome

Customers reaching AWS us-east-1 through Cloudflare faced an outage caused by insufficient congestion management during an unusual high-traffic event. BLOG-2938 1