GitHub unveiled Rubicon, a new Git storage architecture that internal stress tests show can sustain 5,000 concurrent agents and 7,200 pushes per second in a single repository, with read capacity scaling horizontally on demand. The architecture is in limited private preview, with broader rollout planned for early next year, aimed at the sustained high-throughput load created by always-on AI agents.
Horizontal Scaling
水平扩展
Increasing system capacity by adding more server instances rather than upgrading existing ones.
Curated Intelligence
OpenAI introduced Pando, a new Git backend designed to overcome performance bottlenecks in its monorepo caused by the concurrent workloads of Codex agents. Maintaining full compatibility with standard Git protocols, Pando reduced median push latency by nearly 5x and now handles approximately 1 petabyte of Git data daily.
Modal engineers bypassed Kubernetes' strong consistency constraints by building a decentralized sandbox infrastructure that supports one million concurrent sandboxes and tens of thousands of creations per second. The core strategy eliminates O(N) state coordination bottlenecks by treating scheduling as load balancing, making each worker node an independent source of truth and deploying parallel schedulers to avoid etcd performance limits under high pod churn.
The latest MCP specification removes protocol-level sessions, enabling stateless request routing to any instance and significantly simplifying horizontal scaling. This change eliminates the need for sticky sessions and dedicated storage, making serverless architectures like AWS Lambda ideal, but shifts state management responsibilities to the application layer.