OpenAI explains how Pando handles large Git repositories
Pando separates metadata and object storage so capacity can scale independently. During a push, the server advertises a smaller set of refs to reduce negotiation overhead.
Why it matters
When pushes to a large repository slow down, object pack size is only one place to look. Ref negotiation can also add work at the start of a connection. Pando illustrates how to address those costs separately as storage and traffic grow.
For repository operators, start by separating the wait before transfer from the time spent sending objects. Compare small and large pushes in repositories with different numbers of branches and tags. That can help the team see whether most of the time is spent before any data moves. Separating storage also raises an operational question. Teams need to decide how metadata and objects stay consistent when an operation fails midway. That consideration belongs alongside the performance figures when using Pando as a reference for an internal Git service.