· via dev.to (home feed)
Microsoft 365 outage ran 30 hours as health checks drained the Exchange fleet
Microsoft's own incident log puts the November Microsoft 365 outage at 30 hours and 5 minutes, with a latent network bug making health checks pull Exchange servers out of rotation faster than they could recover.

What happened, and for how long
Microsoft 365 was down for 30 hours and 5 minutes if you go by Microsoft's own incident log. A dev.to analysis that rebuilt the timeline from Microsoft's status pages puts the impact window from 10:05 UTC on Thursday, November 12 to 16:10 UTC on Friday, November 13, and points out that the figure in most headlines depends on who counted and when they stopped. The Guardian reported roughly 18 hours, WIRED said 24, and one university status page relaying Microsoft's updates clocked 33 hours and 50 minutes. GeekWire, reading the same log, called it about 30 hours.
The gap comes mostly from the start time. Microsoft's first admin notice went out at 16:17 UTC on Thursday, followed by a wider notice at 18:26 UTC, but the log backdates the start of impact to 10:05 that morning, six hours before anyone was told. Counting to the first signs of recovery yields 18 hours; counting to the final all-clear at 04:17 UTC on Saturday, after the mail backlog was worked through, yields roughly 42. A second dev.to analysis, measuring from when mailboxes stopped opening at about 16:20 UTC, puts it at just under 24 hours, so the sources disagree mainly on where the clock starts.
Microsoft ran the event as two incidents: EX1550543 for Exchange Online and MO1550821 for the services that depend on it, including Teams, SharePoint Online, Copilot, Defender XDR, Intune, Power BI and more, with more than 15 services affected in total. A separate, shorter Teams incident ran from 14:45 to 15:53 UTC the same afternoon; Microsoft has not connected the two. Downdetector counted around 130,000 Outlook reports shortly after 1 p.m. Eastern on Thursday, from 74 countries and territories, with hospitals, airlines, government offices and railway operators among those affected.
How a health check took the system down
Microsoft's preliminary explanation, quoted in the dev.to timeline, is that the incident "was introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change" — a latent bug in the systems that establish secure connections. The company has said it was not a cyberattack and had nothing to do with AI tools.
The other dev.to analysis, by Robert Paschal, reconstructs the mechanism. Exchange servers genuinely could not establish secure connections, so automated health checks correctly reported failure and load balancers pulled the servers out of rotation. But the broken dependency was shared: every server failed the same check at once, and the load balancer, unable to distinguish "this server is broken" from "the thing this server depends on is broken," drained the whole fleet. Every server it removed pushed its connections onto servers already short of capacity, and restarted servers failed the checks again the moment they took traffic back. Client retries piled further load onto whatever remained standing. Microsoft's Charles Lamanna said on X that "cascading failures + thundering herds at our network perimeter" made recovery harder than normal. Microsoft also extended the email timeout window to 48 hours so delayed mail would keep retrying gently rather than bounce.
Second-order casualties
Anthropic's status page logged its own incident for the Microsoft 365 connector, the link that lets Claude read mail, calendar and files, from 16:03 UTC on Thursday to 16:40 UTC on Friday — 24 hours and 37 minutes. Errors on mail, calendar and search were still being reported after Microsoft's own impact window had closed, a dependent service recovering more slowly than the thing it depended on. Anthropic also had an unrelated degradation on Friday afternoon, which the timeline author keeps separate.
The email-only safety alarm
The strangest ripple: at the Hotel Ling Bao in the Phantasialand theme park in Brühl, Germany, a faulty automatic chlorine dosing system over-chlorinated a pool on Friday afternoon. Fifteen people were triaged, and four went to hospital according to the fire brigade's count, though one German paper put the number at five. Per the park's account to broadcaster WDR, relayed by Notebookcheck, the sauna area's maintenance safety system raised its alarm by email only, and with Exchange down the warning never arrived. The fire brigade attributes the release to the dosing-system defect, officials have not confirmed the link to the outage, and the investigation is ongoing.
Why it matters
This is a case study in the false-negative health check, the failure mode where an endpoint correctly reports that a server cannot do its job, but does so across an entire fleet at once because the broken dependency is shared. A load balancer sees only a wall of failing checks and amplifies the outage by draining servers that might have served some requests. The structural fixes are known: probe shared dependencies on an endpoint the load balancer does not use to drain servers, degrade gracefully instead of failing when only one dependency is unreachable, and give retries a budget with backoff and a limit. The timing analysis adds the other lesson: impact began six hours before the first admin notice, and downstream systems — including a pool safety alarm that existed only as an email — recovered on email's schedule rather than their own. Anything safety-critical riding on a store-and-forward system with no delivery guarantee inherits that system's worst day.
- #cloud
- #microsoft-365
- #outage
- #incident-analysis
- #exchange-online