Home / Blog / AWS's US-EAST-1 Outage Shows Identity Can't Be a Single Point of Failure
Engineering

AWS's US-EAST-1 Outage Shows Identity Can't Be a Single Point of Failure

Between 3:05 and 6:21 a.m. PST on November 15, 2026, AWS's US-EAST-1 region broke authentication, sign-ups, and network connectivity for eleven services at once, including AWS IAM Identity Center. AWS has not published a root cause. The part worth sitting with isn't the outage window, it's how many unrelated systems failed the moment identity did.

Just after 3 a.m. Pacific on November 15, AWS engineers were automatically paged for an incident in US-EAST-1, the Virginia region that still anchors a disproportionate share of the internet's backend. By the time AWS's status page caught up, eleven services were listed as impaired: IAM Identity Center, AWS Sign Up, Step Functions, CloudWatch, Amazon Connect, Elastic Load Balancing, OpenSearch Service, NAT Gateway, IoT Device Management, VPC Lattice, and Kiro. AWS's own update described the window plainly: "Between 3:05 AM and 6:10 AM PST we experienced signup errors, authentication issues, and connectivity issues in the US-EAST-1 Region." Authentication was restored by 5:20 a.m., full recovery followed by 6:21 a.m., and a residual Network Load Balancer issue lingered for customers in one Availability Zone cell until AWS recommended they recreate load balancers spanning multiple zones. No root cause has been published, and AWS has said only that it found no evidence of malicious activity.

Identity Center was supposed to be the fix, not the blast radius

AWS IAM Identity Center exists specifically so that organizations can centralize access to the console and to federated applications instead of scattering credentials across dozens of IAM users. That centralization is good security practice right up until the centralized thing goes down, at which point every team that consolidated into it loses access at the same moment. For three hours on November 15, that included anyone trying to sign into the AWS console, anyone whose internal tooling authenticates through Identity Center, and anyone trying to sign up for AWS for the first time. None of those are niche workloads. Account signup failing for three straight hours is the kind of outage that doesn't show up in a postmortem slide about "degraded services," it shows up as a lost prospect who tried a competitor instead.

Engineering teams evaluating custom software built with independent failure isolation instead of maximal reliance on a single managed identity layer aren't rejecting AWS. They're refusing to let one region's bad morning decide whether their own product is reachable.

Eleven services, one region, zero disclosed root cause

The specific list of affected services matters less than what it has in common: nearly all of it sits in the control plane, not the data plane. Step Functions orchestrates workflows. CloudWatch feeds monitoring. NAT Gateway and VPC Lattice move traffic between services. Elastic Load Balancing routes it. When the control-plane layer in one region degrades, the data your application serves might be fine, but your ability to deploy, scale, authenticate, or even observe that it's fine goes dark along with it.

Why "fail over to another region" doesn't save you here

That's a harder failure mode to plan for than a simple storage outage, because the usual mitigation, "fail over to another region," assumes the failing component is regional in the first place. IAM Identity Center is often treated by engineering teams as effectively global infrastructure, the way DNS or a CDN is, precisely because outages in it are rare enough to be forgotten about between incidents.

  • Audit which of your systems treat AWS IAM Identity Center, or any single-region identity provider, as infrastructure that cannot fail. If losing it for three hours would lock your own engineers and customers out simultaneously, that is a single point of failure wearing a managed-service costume.
  • Decouple signup and first-touch authentication from your primary identity dependency where you can. A prospect trying your product for the first time during a vendor outage has no stored session to fall back on, and no patience for your postmortem.
  • Build health checks that independently verify authentication success rates rather than trusting a vendor's status page. AWS's own incident window (3:05 to 6:10 a.m.) was defined only after the fact; teams detecting the spike in auth failures on their own side bought back response time AWS's dashboard did not give them.
  • Treat "no root cause disclosed" as a planning input, not an inconvenience. AWS did not explain why eleven unrelated-looking services failed together, which means teams can't rule out this exact cluster failing the same way again.
The outage lasted three hours and change. The services it took down together are the ones most teams assumed couldn't fail at the same time.

What this changes about build-versus-buy for identity

Nobody should read a three-hour regional incident as a reason to roll a homegrown identity provider instead of using AWS, Okta, or any managed IdP; that would trade a rare outage for a much larger, permanent security liability. What a disclosed-nothing, eleven-service outage should change is how much of your own system's availability gets silently inherited from a single region of a single vendor's control plane. Teams that build custom software with explicit failure isolation around authentication, queued retries for signup, cached session validation, a fallback path that degrades gracefully instead of hard-locking, aren't duplicating AWS's job. They're making sure that when AWS has a bad morning again, their own product has a plan that doesn't depend on AWS publishing one first.

Have a system that has to last?