GitHub pairs immutable Git packs with an atomic reference log
GitHub describes how its new storage design separates immutable Git pack files from repository references. A write first stores the pack durably, then atomically commits the pack manifest and reference changes through its Ref-CAS service. Request-serving replicas cache data locally, while a separate fleet compacts packs, keeping that maintenance work away from machines handling user requests.
GitHub Engineering · Nov 5, 2026
Why it matters
The design shows how a small consistency boundary can coordinate large immutable objects without tying every operation to one repository server.
More context
This follow-up adds the concrete commit sequence and pack-manifest mechanism to GitHub’s earlier migration overview. Reads check reference versions and fetch missing packs into disposable caches. Separate compaction workers reduce competition for request-serving CPU.
Keep in mind
This is GitHub’s engineering report about its architecture and rollout, not an independently measured performance result for arbitrary Git workloads.
Original read