A sudden disruption to corporate email hit governments, banks, and businesses across dozens of countries, while engineers in Redmond struggled in vain to restart servers that crashed again as soon as traffic returned.

Illustrative diagram: the Microsoft 365 failure in three steps, as described by Microsoft.
On November 12, 2026, according to the incident log Microsoft sends to administrators, email issues (Exchange Online) began at 17:17 Italian time; two hours later, at 19:26, Microsoft opened a second one, affecting multiple services in the suite[1]. A domino effect triggered disruptions across several Microsoft 365 tools, Teams, and even connectors used by third-party AI platforms such as OpenAI and Anthropic. Microsoft 365 counts more than 430 million users, according to Microsoft figures cited in the press[2]. On Downdetector, which tracks crowd-sourced user reports, issues reported for Outlook peaked above 130,000 in the United States[1]. In Italy, reports peaked at 19:17: 1,080[3]. Overall, according to Downdetector, disruptions affected users across 74 countries and territories.
Over the following hours, as Japanese government offices resorted to paper records, Philippine courts faced delays, and British schools asked parents to call instead of emailing, one question inevitably arose: how can a global tech giant struggle for more than sixteen hours[4]? And above all, does relying on SaaS (Software as a Service—applications delivered over the cloud without local installations) still make sense?
At 12:00 UTC on November 13, Microsoft reported encouraging signs of recovery in Europe, with service availability and email delivery improving[5], while Japan appeared comparatively more stable and North America was also improving, despite ongoing intermittent issues. The company will publish a full Post-Incident Report with a root cause analysis once mitigation is complete.
Contents
- What broke inside Redmond's systems
- The journey of a request: where Microsoft 365 stalled
- Vendor responsibility: numbers, penalties, and promises
- How to monitor a giant: the four golden signals and the USE method
- Does SaaS still make sense over an on-premises server?
- How we manage infrastructure and reliability
- Continuity is never an accident
What broke inside Redmond's systems
In its initial updates on official status channels, Microsoft pointed to elevated load across its infrastructure in multiple regions. Shortly afterward came the clarification: the issue was not caused by a spike in overall traffic, but rather by a large number of servers unexpectedly becoming unavailable. Ordinary traffic was consequently forced onto the remaining operational machines, putting them under severe strain.
In its Site Reliability Engineering (SRE) handbook, Google describes this exact cascading failure: when some nodes fail, requests shift to the surviving capacity, which eventually runs out of CPU and memory and crashes in turn[6]. According to Microsoft's preliminary findings, the issue appeared to stem from a specific component on the servers responsible for processing automated health checks—the periodic probes a system uses to verify whether a server is healthy and ready to accept traffic. As a precaution, the company paused all code deployments (scheduled update releases) and configuration changes, and began rebooting all affected servers.
Recovery efforts hit a double bottleneck, however. On one hand, Microsoft observed that servers returned to service tended to become unresponsive again as soon as they received traffic. This phenomenon is also well documented in SRE literature: after an outage, a handful of rebooted servers must shoulder the full backlog of demand with reduced capacity, causing them to collapse again. On the other hand, engineers identified that a process handling secure connections was becoming overwhelmed, halting request processing altogether. To stabilize the traffic flow, Microsoft applied selective traffic restrictions, validated an alternative health check method, and began reverting a recent network configuration change that correlated with the start of the incident.
Charles Lamanna, Microsoft executive vice president for Copilot, Agents and Platform, confirmed publicly that there was no indication of malicious activity or a cyberattack. To prevent delayed messages from bouncing, Microsoft extended the email timeout window—the maximum duration the system waits before treating a message as undeliverable—to 48 hours.
The journey of a request: where Microsoft 365 stalled
To understand where such an intricate chain can break, let us follow the path taken by a single request: a business manager opening Outlook on their computer at eight in the morning.

Illustrative diagram: the lifecycle of a request in a service like Microsoft 365. Highlighted in orange are server health checks, where Microsoft suspects the breakdown began.
The application first queries DNS (Domain Name System, the internet's phone book that translates human-readable addresses into numerical endpoints). Once resolved, the request reaches the CDN (Content Delivery Network—a globally distributed network of edge servers close to users that also acts as a shield against attacks). This highlights the cloud's primary advantage, scalability: delivering services worldwide without shipping hardware or installing software on premises. Once inside the perimeter, the connection reaches the load balancer, the component that distributes incoming calls and continually probes nodes via health checks so work is sent only to instances ready to process it.
From the load balancer, the request passes to the application servers—the machines running the actual software. In cloud architectures, these nodes scale dynamically, adding or removing instances as demand fluctuates. This is where multi-tenancy operates: a single massive infrastructure hosts a great many companies (tenants), logically isolated from one another. This shared model introduces the "noisy neighbor" problem: if one tenant triggers excessive workloads, it risks saturating server resources and degrading performance for others. To protect infrastructure, Microsoft's documentation details how throttling is enforced, rejecting excess requests with HTTP status code 429 ("Too Many Requests").
Past the application layer, the request reaches mailboxes and databases, replicated across multiple instances; if write demand outpaces real-time throughput, messages sit in a processing queue. The entire system is updated continuously via deployment rings, rolling out patches and feature updates without manual user intervention. The flip side is that any change can conceal unforeseen regressions. Meanwhile, every individual request emits telemetry data that informs software optimization, pricing tiers, and technical support.
Our multi-tenant SaaS infrastructure for Odoo also follows the same logical sequence. In the November 12 disruption, according to Microsoft's initial statements, the weak point may have been the health check component itself: with the checks failing, many servers were taken out of service and their load shifted onto the remaining machines, putting them under strain.
Vendor responsibility: numbers, penalties, and promises
A cloud application provider shoulders full responsibility for reliability, security, and operational compliance. When infrastructure fails, the fallout impacts both brand reputation and customer contracts.
The service commitment is governed by an SLA (Service Level Agreement), centered around uptime (the percentage of time the service is available and operational). Under Exchange Online terms, Microsoft allows customers to claim a service credit, subject to Microsoft approval: 25% of the monthly fee if availability drops below 99.9%, 50% below 99%, and 100% below 95%[7].
However, Microsoft's agreement explicitly states that service credits are the customer's sole and exclusive remedy for any outage: no credits or compensation are provided for lost revenue, operational disruption, or indirect business damages.
To grasp what these percentages mean in practice, one only needs to look at the Google SRE availability table: a 99.9% uptime target allows for no more than 43.2 minutes of cumulative downtime in a full month, or 8.76 hours per year[8]. Against these thresholds, November's outage exceeded sixteen hours.
How to monitor a giant: the four golden signals and the USE method
Maintaining visibility across complex infrastructure processing millions of requests requires rigorous discipline. Google's SRE handbook, also highlighted in Grafana Labs documentation[9], defines the "four golden signals" that must be continuously monitored:

Illustrative diagram: the four golden signals on a monitoring dashboard during an outage where traffic stays flat but saturation, error rates, and latency surge.
- Latency. The time it takes to service a request. Tracking only mathematical averages is deceptive: as Google's handbook notes, if a service has an average latency of 100 milliseconds across 1,000 requests per second, 1% of those requests can easily take 5 seconds[10]. This is why monitoring focuses on percentiles such as p95 (the threshold within which 95% of requests resolve) or p99, exposing what the slowest slice of traffic experiences.
- Traffic. The measure of overall demand, quantified in requests per second (RPS) or concurrent sessions.
- Errors. The rate of failed requests, distinguishing explicit HTTP 5xx codes (server failures) from 4xx codes (invalid or rate-limited requests).
- Saturation. How close the system is to capacity: CPU utilization, memory allocation, and database connection pool limits.
In multi-tenant environments, these metrics are complemented by per-tenant instrumentation: resources consumed by each organization, HTTP 429 rate-limiting responses, and API call volumes per account. At the operating system and bare-metal layer, systems follow Brendan Gregg's USE Method (Utilization, Saturation, Errors[11]): checking the percentage of time resources are busy, pending work queues, and hardware or network errors. On databases, key checks track active versus pool-limit connections, slow query counts, and cache hit ratios.
This data feeds operational dashboards. At our edge, for instance, a single real-time console displays requests per second (broken down by status code and per tenant), overall response times across p90, p95, and p99 alongside per-tenant p95, aggregate bandwidth, and active connections. Alerts trigger on user-facing symptoms rather than infrastructure noise, enabling rapid intervention without alert fatigue. Viewed through this lens, the Microsoft outage is textbook: external demand had not spiked, but effective serving capacity vanished; the remaining servers became saturated, leading to dropped connections and expired requests.
Does SaaS still make sense over an on-premises server?
Faced with failures of this magnitude, it is natural to wonder whether keeping software on a physical machine in the office was safer all along. In our view, the answer still clearly favors the cloud. A comparison with legacy systems (old software installed on premises, often isolated and out of date) highlights operational advantages that a single incident does not erase:
- Continuous updates and security. Security patches (vulnerability fixes) and software upgrades are handled centrally by the vendor without disruptive local maintenance.
- Geographic redundancy. Cloud providers replicate workloads across multiple physical hosts and separate data centers; a local office server can bring operations to a standstill during a power outage or hardware failure.
- Dedicated engineering. Very few small and medium-sized businesses can afford an in-house operations team monitoring latency percentiles, saturation limits, and queue depths around the clock.
- Elastic scalability. When headcount grows, computing resources adjust without procuring and configuring new physical hardware.
That does not rule out the risks of relying on a single provider. To protect itself, every business can take a few concrete precautions:
- Map process dependencies. Clearly identify which operational workflows stall when email or collaboration platforms go down.
- Establish an independent fallback channel. Keep a secondary communication route (such as a separate messaging tool or standard phone lines) configured in advance and hosted off the primary vendor's infrastructure—because a backup sharing the same core systems will fail at the exact same time.
- Review service commitments. Examine your vendor's uptime SLA terms, incident communication policies, and disaster recovery commitments.
How we manage infrastructure and reliability
At QTNOVA, we manage scalable multi-tenant SaaS infrastructure for clients running Odoo ERP. We share direct responsibility for ensuring our platforms remain accessible and dependable, which is why we apply these identical operational principles across our own stack.
We measure latency, traffic, and error rates on every inbound request, isolating metrics per client. We track saturation across host nodes, application containers, database connection pools, and background workers. A capacity alert tells us the load is rising before the servers still running can no longer absorb the loss of one of them: that is exactly the scenario of the Microsoft outage, and we want to know before it happens. For each business, we independently track disk usage, active user counts, and any emerging HTTP 500 errors.
Alert routing matches operational urgency: critical anomalies are sent via email and via a notification on the engineers' phones. Each alert automatically opens a ticket in our helpdesk, classified by automated triage before an engineer intervenes. The monitoring system itself is watched for any hang-ups and configured with the same priority as production services, ensuring it is not the first victim of a crisis it is meant to signal.
For databases, we run a standby replica ready to failover immediately, paired with dedicated alerting should replication lag drift. Backups take two isolated pathways to geographically redundant object storage with a third-party provider. All traffic passes through a content delivery network (CDN) that filters automated scans before they touch origin servers. Every infrastructure change follows a documented runbook with an explicit rollback plan, tested in staging and executed step-by-step by an engineer. For managed clients, version upgrades and security updates are included in the subscription fee, with no operational downtime and the final cutover scheduled outside business hours.
This same monitoring infrastructure caught the first Outlook connection errors on the evening of November 12. In our view, what matters most for anyone working on a managed platform is that everything runs smoothly, just as people expect from the services of the big tech companies; unexpected problems can happen to anyone, but the difference lies in noticing them first and managing the event.
Continuity is never an accident
No IT system in the world is entirely immune to failure—even the largest technology providers experience unplanned outages. The real difference between suffering through extended downtime and safeguarding business continuity lies in end-to-end visibility, containing incidents before they spread, and the speed of the engineering team watching over the systems.
If your email or ERP stopped answering tomorrow morning, does your organization have a plan to keep working? If you want to evaluate how dependent your business is on a single vendor, let's talk.
Want to know more? Leave us your email below and we'll get back to you as soon as possible.
Sources:
- la Repubblica — “Microsoft, down mondiale di Outlook e Microsoft 365. Disagi per aziende e governi, cosa è successo” (November 13, 2026) – repubblica.it
- BBC News — “Schools, GPs and businesses hit by global Microsoft outage as company scrambles to solve issue” (November 13, 2026) – bbc.co.uk
- Fanpage.it — “Microsoft 365 e Outlook down oggi, problemi con le app: cosa sta succedendo” (November 12, 2026) – fanpage.it
- ANSA — “Outlook down, Microsoft: 'Servizi in miglioramento, ripristino a fasi'” (November 13, 2026) – ansa.it
- Microsoft — “Users may be unable to connect to multiple Microsoft 365 services” (November 13, 2026) – status.cloud.microsoft
- Google — “Site Reliability Engineering: Addressing Cascading Failures” (2017) – sre.google
- Microsoft — “Service Level Agreement for Microsoft Online Services” (October 1, 2026) – microsoft.com
- Google — “Site Reliability Engineering: Availability Table” (2017) – sre.google
- Grafana Labs — “Grafana dashboard best practices” – grafana.com
- Google — “Site Reliability Engineering: Monitoring Distributed Systems” (2017) – sre.google
- Brendan Gregg — “The USE Method” (August 24, 2017) – brendangregg.com