Cases2 cases

1 to 2 of 2

1. IL-0124 Capacity changes queued while a circuit breaker was engaged ran all at once when it was released, taking multiple OpenAI inference engines out of service and raising API error rates

   OpenAI2026

   Symptoms

   Between about 4:26 PM and 4:56 PM PST on 4 March 2026, some OpenAI API requests to several models failed or took longer than usual, and some Codex traffic was affected too. The overall API success rate dropped temporarily.

   High error rateHigh latency

   Trigger

   A protective circuit breaker that had paused some scheduling operations was released, and the infrastructure actions queued during the pause were executed at the same time.

   Root cause

   Software defect

   Component

   Automated capacity manager

   Contributing factor

   Automation amplificationChange with a delayed effectNo safeguard on destructive changeMonitoring blind spot

   Remediation

   • manually reassigned infrastructure resources to the affected models
   • brought the removed inference engines back online and reassigned them until error rates returned to normal

   Action items

   • Improve safeguards in the automated capacity management systems so they detect and respond to state drift
   • Improve observability of circuit breakers and scheduling systems so their status and effects are easier to identify
   • Add safeguards so queued actions cannot run in large batches after long pauses
   • Improve monitoring and alerting around capacity reconciliation problems

   Lessons

   Pausing one automated system does not stop the systems that feed it. Changes queued during the pause reflect the state at the time they were queued, and releasing the queue all at once applies those stale decisions together. The fixes OpenAI names are detecting state drift, making circuit breaker status visible, and preventing large batches from executing after a long pause.

   View case
2. IL-0118 A global change to Kubernetes namespace labels made the networking layer recompute metadata and overwhelmed the control planes of three of OpenAI's largest GPU clusters

   OpenAI2024

   Symptoms

   From 10:15 am PT on 25 November 2024 a large share of both API and ChatGPT traffic failed with timeouts and 503 errors.

   High error rateHigh latencyControl plane unresponsive

   Trigger

   A global change to Kubernetes namespace labels caused the networking layer to recompute metadata. The change had rolled out to the small clusters without problems.

   Root cause

   Configuration change

   Component

   Kubernetes API server

   Contributing factor

   Excessive blast radiusLimited debug access

   Remediation

   • moved workloads out of the impacted clusters by hand
   • with the cloud provider, scaled up the number of control plane nodes while scaling down the number of data plane nodes, which let the control planes recover

   Action items

   • Lock namespace changes so this incident cannot recur
   • Freeze certain types of infrastructure deploys
   • Allowlist identities to prevent a recurrence of this class of incidents
   • Improve the GPU load balancing logic so traffic is re-routed automatically and more effectively when possible
   • Improve change management for fleet-wide infrastructure changes

   View case

By root cause2