20× the CI traffic without getting slower: How we rebuilt Git serving at Datadog
Summary
Datadog explains how they built gitretriever to serve CI Git traffic at scale. The system uses independent mirrors and relays to avoid CPU bottlenecks on a single backend, achieving ~100 million requests per week with median latency around 40 ms, despite ~20× traffic growth. Key design choices include distributing pulls across branches, reusing packfiles via a pack cache, and a small read-only API to answer common queries without cloning.