US - Document Understanding and IXP - Degraded Performance

Update timeline

1. investigating Sep 01, 2026, 01:49 PM UTC

   We are investigating reports of degraded performance impacting digitization and extraction for Document Understanding and IXP in US. Impact: Users may experience slowness and failed Document Understanding runtime operations. Our teams are working to identify the cause and will share more details as the investigation progresses.

2. monitoring Sep 01, 2026, 02:07 PM UTC

   We have identified the issue and applied a mitigation. Services are returned back to healthy state and we are monitoring the services.

3. resolved Sep 01, 2026, 03:06 PM UTC

   This issue has been fully resolved and services are stable now. We'll post more details of the incident to status page soon.

4. postmortem Sep 15, 2026, 11:08 AM UTC

   ## Customer impact On September 1, 2026, between 10:00 UTC and 15:06 UTC, customers using Document Understanding in the US region experienced severe delays in document digitization, the step that converts uploaded files into machine-readable text. Document Understanding and IXP operations that depend on digitization were affected in the same way. Digitization times increased substantially compared to normal, and a backlog of 15,000 to 20,000 queued jobs built up before being fully processed. ## Root cause The issue was caused by a significant increase in digitization traffic in the US region, which exceeded the capacity of the database that distributes digitization work to processing workers. This database had operated reliably at its existing size for several years; the recent growth in traffic pushed it beyond its limits. Once saturated, the database could hand out work far more slowly than incoming demand required, so a backlog accumulated for as long as demand exceeded that rate. ## Detection Automated internal checks on a dependent service detected the slowdown, and the incident was opened at 13:33 UTC. ## Response Our engineering team identified the work-distribution database as the cause by 13:33 UTC and increased its capacity at 13:40 UTC. As a consequence, the backlog cleared between 13:45 and 14:00 UTC, impact was mitigated at 14:07 UTC, and processing times were back to normal at 15:06 UTC. The service has operated normally since. ## Follow up 1. We have increased the capacity of the affected database and are reviewing capacity headroom across all databases supporting this service, in all regions, against current and projected volume. 2. We are tuning our automated alerting so that sustained high load and processing delays are detected and escalated earlier, before they become customer-visible. 3. We are reviewing and optimizing the indexes on the databases supporting this service so that queries remain efficient as data volume grows.