GitHub outage exposes flaws in autoscaling and retry systems
VS Code bug amplified traffic
GitHub has attributed a nearly eight-hour outage this week to a combination of saturated load balancers, a faulty autoscaling policy and a software bug in Microsoft's Visual Studio Code that dramatically increased traffic.
The disruption began at 13:28 UTC on 17th August and was not fully resolved until 21:15 UTC, lasting seven hours and 47 minutes.
Users experienced elevated error rates across services including Issues, APIs, Pull Requests, Actions and Copilot.
GitHub said the immediate problem was network saturation affecting load balancers at its Central US facility.
The disruption began when an Istio sidecar, a component used to manage network traffic between services, reached its concurrency limit.
Under normal circumstances, GitHub's autoscaling system should have responded by adding capacity.
However, the company said a misconfigured autoscaling policy was monitoring the host service rather than the sidecar's concurrency limit. As a result, additional capacity was not automatically added when it was needed.
The situation was then worsened by retry behaviour that generated more traffic and placed further pressure on internal load balancers.
Engineers reduced gateway retries through a temporary code change and configured load balancers to reject incoming Copilot Token Service requests with HTTP 403 responses.
VS Code bug amplified traffic
GitHub said delayed responses from a single internal endpoint triggered a previously unidentified retry bug in Visual Studio Code.
The company said the behaviour amplified Copilot token traffic by about 10 times, contributing to a prolonged recovery for the Copilot Token Service.
Most GitHub services had recovered by 16:36 UTC, while Actions returned to normal at 18:03 UTC.
The Copilot Token Service took considerably longer, with recovery completed at about 21:02 UTC.
GitHub also said several scraping attacks targeting its codeload endpoints complicated efforts to restore normal service.
The company has promised to correct its autoscaling policies, review retry limits and audit Istio concurrency settings. It also plans to address the Visual Studio Code behaviour that amplified traffic to the Copilot service.
Outage comes as coding rival launches
The disruption comes as competition in the software development market intensifies.
AI coding company Cursor has launched Origin, a code-hosting platform offering features similar to those developers use on GitHub, including collaborative coding, repository management and pull requests.
Cursor says Origin is designed to work alongside GitHub rather than replace it. Developers can connect their GitHub organisations, import repositories and move code between the two platforms.
The launch gives developers another option at a time when reliability and availability are increasingly important to software teams that depend on online development infrastructure.
GitHub remains the world's largest source-code hosting platform. The company said last year that it had about 180 million developers using its service. Founded in 2007, GitHub was acquired by Microsoft in 2012.
The latest incident is not the company's first major service disruption.
GitHub experienced a global outage in February 2022, followed by further failures in March that year and May 2023.
A major outage in January 2025 also affected services including pull requests, issue tracking and repository access. GitHub said that incident was caused by a failure in its caching infrastructure.