Microsoft’s outage report shows how failover helped spread a gateway failure
The preliminary account links a latent defect, shifted traffic and reduced capacity. Recovery routing became part of the load that overwhelmed the fleet.
Microsoft’s preliminary account of the November 12–13 Microsoft 365 outage describes an uncomfortable failure mode: the mechanisms that moved traffic away from unhealthy servers helped overwhelm the remaining ones. A latent software defect, a change in traffic and reduced available capacity combined at the network gateway used by Exchange Online and other services.
The incident is resolved, so this is an explanation of the failure rather than a claim that Microsoft 365 is currently unavailable. The important question is why a service designed to route around trouble could spread it instead.
Three conditions lined up at the gateway
Microsoft’s preliminary report, reproduced by Office Watch, identifies a defect introduced in July in code specific to its networking gateway servers. Under certain loads, contention inside a Windows process exhausted resources. The company says testing missed the problem because the lab’s traffic mix and lower-core-count hardware did not reproduce the conditions that exposed it.
Two circumstances increased the pressure: Teams traffic had moved to the primary gateway, while an unrelated issue had left nearly 14% of gateway capacity out of service. Microsoft says that reduction was not enough to cause trouble on its own. The failure depended on the combination.
That is a more useful explanation than assigning blame to one large traffic number. Capacity is not simply a fixed quantity of requests a machine can handle. The kind of work and the way concurrent operations interact can change the effective limit. A system can have apparently adequate headroom while a specific code path becomes the constraint.
Redirecting traffic can spread the same problem
Ordinary failover assumes that a healthy destination can take work from an unhealthy one. If the same latent defect exists across the fleet, the transferred load can push the destination into the same failure state. The act of avoiding one broken server then reduces the number of remaining healthy servers.
The preliminary account says more than 80% of the gateway servers became unresponsive at the worst point. It also says the normal partitions intended to contain failure did not stop the spread. That second point deserves attention: redundancy is less protective when apparently separate capacity shares the same trigger and containment does not hold.
This does not make failover a bad design principle. It shows why a recovery path needs to be tested as a workload in its own right. The traffic that arrives during a failure can differ from the steady traffic used to validate normal operation.
Restarting was not enough
The report describes returning servers being overwhelmed by waiting traffic, and an older defect that ordinary recent-change rollback could not quickly remove. Engineers eventually applied a targeted correction and restored groups of servers progressively. Microsoft’s public incident updates, archived by the University of Pennsylvania, record the restoration and subsequent review status.
For customers, the distinction is practical. A service-side connection failure is not evidence that every affected laptop needs repair. Nor does a successful sign-in prove that every delayed business process has completed. Recovery of access and recovery of accumulated work are separate observations.
The broader infrastructure lesson is about what counts as a realistic test. Testing only ordinary load on smaller machines can miss the interaction between newer hardware, altered traffic and a degraded fleet. Microsoft’s final review is expected to supply further detail; the preliminary account already shows why spare capacity alone cannot guarantee a safe escape route.