Moving 240 TB of takes to the v3 bucket layout

2026-10-21 · 6 min read

We finished the storage migration this week. Dual-read ran for two weeks; the cutover itself took 47 minutes and one bad assumption about eventual consistency in the manifest index.

What went wrong

The manifest index was written before object headers settled for objects over 4 GB, so 312 restores returned stale checksums. The fix was to re-derive manifests from object bytes rather than headers.

Status and any follow-ups: service status. API surface is unchanged — see storage docs.