Customer impact

Between February 24, 2026 at 5:21 pm UTC and 6:19 pm UTC, there was a partial outage affecting Document Understanding in the US region. During this window, some customers were unable to classify or extract documents—both out-of-the-box and custom models—with requests timing out or returning errors. We understand how disruptive this is when document processing is part of your critical workflows, and we're sorry for the impact.

Root cause

The issue stemmed from resource exhaustion on shared compute infrastructure that supports Document Understanding. A surge in request volume coincided with degraded behavior from a model in the processing pipeline, and together these consumed more capacity than was available. The resulting backlog caused incoming classification and extraction requests to time out or fail.

We identified the source of the excess load quickly and took steps to reduce it, allowing the service to recover within approximately one hour.

Detection

Our automated monitoring triggered alerts at 5:26 pm UTC—five minutes after the first signs of elevated failures in the US region. The engineering team quickly identified the affected endpoints and confirmed that the root cause was compute resource exhaustion.

Response

Within minutes of detection, we identified the source of the excess load and reduced it by adjusting traffic routing to the affected Document Understanding endpoints. We continued monitoring closely until error rates returned to baseline, and full recovery was confirmed by 6:19 pm UTC.

Follow up

We're taking several concrete steps to prevent this type of issue from recurring:

• We are completing infrastructure isolation work so that different categories of document processing models run on dedicated compute resources, reducing the risk of one workload affecting another.
• We are adding capacity-aware safeguards to protect production services during unexpected traffic spikes.
• We are improving our monitoring and dashboarding to detect workloads consuming disproportionate resources before they impact service availability.