20× the CI traffic without getting slower: How we rebuilt Git serving at Datadog

Why we picked it · the editor's summary

Datadog's CI Git backend was buckling as AI coding agents drove an order-of-magnitude rise in fetches. Two engineers built GitRetriever: a small fleet of mirrors that poll GitHub with parallel per-branch fetches, and an autoscaled fleet of relays that hand those packfiles to CI jobs without rebuilding them. About half of pack-building fetches now come straight from cache; traffic grew about 20× in four months while median serve latency stayed around 40 ms, and the old backend's fetch CPU fell three to four times. The migration ran behind feature flags with automatic fallback to the old path, one repository group at a time, largest monorepo first. If your CI clones on every job, copy the read-only API for ref lookups and single-file reads before anything else.