Skip to Main Content

March 2026

9

Watch our 'Azure Incident Retrospective' video about this incident: 

What happened? 

Between 23:20 UTC on 9 March and 19:32 UTC on 10 March 2026, a platform issue resulted in impact to the Azure OpenAI Service. Impacted customers experienced HTTP 400 and HTTP 429 error responses, specifically for the GPT-5.2 model. All other GPT models were unaffected during this time.

This incident impacted customer resources and queries in the following regions: Australia East, Central US, East US 2, Korea Central, Norway East, Sweden Central, and UK South. 

What went wrong and why? 

The Azure OpenAI Service processes customer requests through model engines deployed across multiple Azure regions and supported by traffic routing systems. Depending on the selected deployment model (Global, Data Zone, or Regional), note that requests may be routed across multiple Azure regions within defined geographic boundaries to support availability and resilience, while customer data remains stored at rest in the selected Azure region. Learn more at: .

A recent update to the Azure OpenAI GPT 5.2 model introduced a configuration change that was not compatible with the version of the model engine code running in production. As part of this update, certain feature settings were enabled to improve service efficiency and resilience – however, the deployed engine version did not yet support those settings. As a result, when customer requests were routed to engines with this mismatch, the service was unable to process these requests correctly.

Generally speaking, any updates to the service are rolled out in line with our Safe Deployment Practices (SDP) which deploys to different regions gradually, in stages. During this update, the earlier stages of the rollout did not include sufficient backend model instances for this issue to surface before the update progressed to additional regions. As a result, the rollout had completed its deployment across our fleet before we were able to determine customer impact.

During mitigation of the primary issue, we identified a secondary issue that affected service recovery. Azure OpenAI relies on internal telemetry to understand real-time service capacity across regions and to route traffic accordingly. At the time recovery actions were underway, an unrelated issue in this internal telemetry system led to incomplete capacity information being leveraged. As a result, traffic routes were temporarily being determined using incomplete data, which led to a disproportionate amount of traffic being directed to a limited set of available regions. This created additional resource pressure in those regions and resulted in continued intermittent request failures (HTTP 429 errors) for some customers, even as the rollback of the configuration issue was progressing and other regions were actually available to receive requests. Once the routing updates were successfully completed and full capacity information was restored across regions, traffic distribution normalized and service recovery progressed as expected.

During the incident, we also identified a monitoring gap related to anomalous HTTP 400 error patterns. While HTTP 400 responses do occur during normal service usage, our monitoring was not configured for service‑side anomalies, only for client-side errors – which are typically caused by incorrect parameters in user requests. This miss in monitoring during the initial stages of the incident delayed detection and response.  

How did we respond? 

  • 23:20 UTC on 09 March 2026 – Customer impact began, triggered by the recent service update. 
  • 00:19 UTC on 10 March 2026 – We detected the issue via service monitoring. This prompted us to begin our investigation, engage with other teams to troubleshoot, and start developing a hot fix. 
  • 03:18 UTC on 10 March 2026 – We determined that a rollback could mitigate more quickly than hotfix, so started the rollback for the impacted model.
  • 10:55 UTC on 10 March 2026 – Traffic routes were determined to be using incomplete data, due to the aforementioned dependency issue. 
  • 12:40 UTC on 10 March 2026 – We identified and investigated resource constraints.
  • 18:00 UTC on 10 March 2026 – Rollback actions completed across all affected regions. 
  • 19:30 UTC on 10 March 2026 – Full capacity information was restored across regions, traffic distribution normalized. 
  • 19:32 UTC on 10 March 2026 – Once monitoring confirmed stable recovery, we determined the service was fully restored and all customer impact had been mitigated. 

How are we making incidents like this less likely or less impactful? 

  • We have already conducted additional engineer training on our operating procedures – including which scenarios can be quickly rolled back – to reduce the time to mitigate similar issues. (Completed) 
  • We have incorporated storing a ‘cached’ known good version of the traffic routing details, as an additional layer of resilience in case the dependent service is not able to serve the latest information on regional capacity availability. (Completed) 
  • We have improved our monitoring surrounding HTTP 400 errors, by establishing thresholds of errors on the service side. (Completed) 
  • To expand that monitoring further, we are improving our anomaly detection surrounding HTTP 4xx errors – to alert on anomalous error rates that may not meet our usual thresholds. (Estimated completion: April 2026) 
  • We are incorporating additional signals to our deployment systems, to reduce potential impact by stopping problematic rollouts automatically. (Estimated completion: May 2026)
  • Finally, we are improving our safe deployment practices by ensuring that early stages have sufficient backend model instances to catch issues like this earlier. (Estimated completion: June 2026) 

How can customers make incidents like this less impactful?

  • Consider reviewing our guidance and best practices related to Business Continuity and Disaster Recovery (BCDR) scenarios for Azure OpenAI:
  • Consider reviewing and implementing our best practices surrounding retry patterns, especially with exponential backoff, to improve workload resiliency during intermittent issues:
  • More generally, consider evaluating the reliability of your applications using guidance from the Azure Well-Architected Framework and its interactive Well-Architected Review:
  • The impact times above represent the full incident duration, so are not specific to any individual customer. Actual impact to service availability varied between customers and resources – for guidance on implementing monitoring to understand granular impact:
  • Finally, consider ensuring that the right people in your organization will be notified about any future service issues – by configuring Azure Service Health alerts. These can trigger emails, SMS, push notifications, webhooks, and more:

How can we make our incident communications more useful?

You can rate this PIR and provide any feedback using our quick 3-question survey:

< View full status history