Incident with Actions

Final update

On May 20, 2026, between 16:00 UTC and 17:45 UTC, GitHub Actions customers experienced run start delays exceeding 5 minutes. Approximately 4.5% of all runs were delayed during the impact window, with scale set jobs disproportionately affected. 30% of scale set jobs were delayed and 4% failed to start entirely. The incident was caused by a misconfigured health check on an internal service that assigns jobs to runners. A brief latency spike in an upstream dependency triggered health check failures across several pods, removing them from service and concentrating load on the remaining capacity. The added load drove memory pressure that escalated into a cascading failure in one regional cluster, leaving it unable to self-recover. Responders mitigated the incident by scaling capacity in the healthy regional clusters and draining traffic away from the impaired one, after which run start latency recovered. To prevent recurrence, we are strengthening our health check configuration to avoid cascading failure scenarios and evaluating automated mitigations to rebalance traffic when a region is degraded.

1. Resolved · May 20, 20:14 UTC
   On May 20, 2026, between 16:00 UTC and 17:45 UTC, GitHub Actions customers experienced run start delays exceeding 5 minutes. Approximately 4.5% of all runs were delayed during the impact window, with scale set jobs disproportionately affected. 30% of scale set jobs were delayed and 4% failed to start entirely. The incident was caused by a misconfigured health check on an internal service that assigns jobs to runners. A brief latency spike in an upstream dependency triggered health check failures across several pods, removing them from service and concentrating load on the remaining capacity. The added load drove memory pressure that escalated into a cascading failure in one regional cluster, leaving it unable to self-recover. Responders mitigated the incident by scaling capacity in the healthy regional clusters and draining traffic away from the impaired one, after which run start latency recovered. To prevent recurrence, we are strengthening our health check configuration to avoid cascading failure scenarios and evaluating automated mitigations to rebalance traffic when a region is degraded.