Explains how AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →-scale Git traffic and data center power limits drove a storage redesign. Useful for anyone running agents that hammer repositories.
Key takeaways · AI-distilled
GitHub says the slowest healthy replica set write latency for a whole repository, because its Spokes quorum waited on every healthy replica while holding a repository-wide lock. A 25ms cross-region round trip therefore capped throughput for the busiest repos.
The fix narrows the quorum from the whole repository to individual Git references. Writes to different branches, such as agents each pushing their own, no longer queue behind each other; GitHub reports 3-4x higher throughput in internal tests with realistic traffic.
A "virtual write replica" makes Azure Blob Storage a quorum member, so a write is acknowledged only once durable in object storage, and it batches writes on disjoint references into one round trip. Read replicas can then scale out on demand without joining the quorum.
GitHub says reads outnumber writes 100 to 1,000 times in most repositories, so decoupling them lets each get its own memory and compute. It names write throughput as the next limit, since Git objects are content-addressed and only a small part of each write needs ordering.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
