GitHub已将99.99999%的代码库副本迁移至Microsoft Azure云平台，与自有数据中心形成并行存储1。第一个代码库副本于2025年8月抵达Azure中央美国地区1。该迁移采用了按引用序列化和虚拟副本等新型技术方案，旨在突破传统架构在地理分散场景中的性能瓶颈。
剩余0.00001%的超高频访问代码库因受约25毫秒跨地域往返延迟限制而仍在迁移中1。为解决这一瓶颈，GitHub引入了按引用的仲裁机制，使不同引用的写入操作得以并发执行，内部测试显示写入吞吐量提升3至4倍1。虚拟副本充当对象存储网关，无需维护持久状态即可作为仲裁成员参与分布式共识1。每个代码库拥有5个副本，Azure目前服务约50%的Git流量1。虚拟读副本在内部分布式负载测试中突破了现有架构的性能上限1。通过解耦存储与计算，该架构使读操作能够独立扩展，为代理规模级开发工作流奠定基础设施基础1。
GitHub has completed the migration of 99.99999% of its code repositories to Microsoft Azure, introducing architectural innovations designed to overcome geographical latency constraints and enable infrastructure capable of supporting agent-scale development 1. The first repository replica arrived at Azure's Central US region in August 2025, with each repository now maintained across five copies, of which Azure currently serves approximately 50% of Git traffic 1.
The remaining 0.00001% of repositories remain constrained by approximately 25 milliseconds of cross-region round-trip latency 1. To address write throughput limitations imposed by this latency, GitHub has implemented a reference-based serialization scheme that allows write operations on different references to proceed concurrently, with internal testing demonstrating a 3-4x increase in throughput 1. The infrastructure employs virtual replicas that function as object storage gateways without requiring persistent state, enabling them to participate as arbitration members in the distributed system 1. In internal distributed load testing, virtual read replicas have broken through the performance ceilings of the existing architecture, while decoupling storage from compute enables independent scaling of read operations 1.
评论
还没有评论，欢迎留下第一条。