AIBRIEF
GitHub was offline for nearly eight hours on August 17 after its autoscaler refused to add capacity and a VS Code bug re sent failed requests, multiplying traffic about tenfold.
GitHub said its August 17 outage was a feedback loop between two of its own systems, in a post-mortem published Tuesday.
An autoscaler in GitHub's Central US data center, the system that adds server capacity on demand, hit its concurrency ceiling and refused to scale up, according to GitHub's status page. Slow replies from that pod tripped a long-dormant bug in the VS Code editor that re-sent failed requests without backoff, amplifying GitHub's incoming traffic by roughly ten times. The autoscaler would not add capacity; the retry bug would not stop asking.
The amplification showed up in the Copilot Token Service, which jumped from a baseline of 7,000 to 9,000 requests per second to 70,000 to 100,000 during the storm, according to a post-mortem analysis. The flood starved GitHub's login and authentication path, taking down Issues, Pull Requests, Actions, and Copilot. Peak error rates reached roughly 20% on web and API traffic and 50% on archive and raw-content downloads, The Register reported.
Most services recovered by 16:36 UTC; Actions stayed degraded until about 18:03, and Copilot was fully restored by 21:02. GitHub said it "must accelerate" reliability work already underway since March. The lesson, per the same post-mortem analysis: client retries need jitter, ceilings, and circuit-breakers, and autoscalers need to fail without inviting coordinated client storms.