Warehouse Data Recovery After a Measurement Layer Failure
A failed warehouse change made 71 days of AI crawler data unreliable. Because the raw server logs survived, the measurement layer could be rebuilt without rewriting the original evidence.
The recovery separated evidence from interpretation. Raw server requests remained the source. Classifications, summaries, query associations, cache rows, and dashboard values were treated as products that could be rebuilt.
The retained window covered 36 dates from August 5 through September 9. Independent aggregation matched 153,231 usable crawler requests across 4,150 cached site and day groups, with zero count or purpose-split differences.
Historical recovery covered 35 dates from June 29 through August 2. It reconciled 3,751,275 source requests across 3,027 site and day groups and produced 233,663 eligible crawler requests under the September 9 classification rules.
August 3 and August 4 remain unavailable. They were not converted into zeroes.
The apply process replaced 280 related table and day states atomically. All 280 matched the reviewed plan. The 6,621,793 live raw request records remained unchanged.
Evidence
- 36 retained dates rebuilt from August 5 through September 9 - 153,231 retained-window usable crawler requests matched - 0 count or purpose-split differences - 35 historical dates recovered from June 29 through August 2 - 3,751,275 historical source requests reconciled - 233,663 historical eligible crawler requests restored - 71 validated dates in the cache ledger - 2 dates, August 3 and August 4, preserved as unavailable - 280 table and day replacements matched the plan - 6,621,793 live raw request rows unchanged - 33 current Performance accuracy regression checks
Methodology
The recovery used live raw requests for the retained window and isolated current and archived log evidence for older dates. The September 9 classification and vendor ranges were applied retrospectively and recorded as provenance.
Proposals were prepared outside the published tables, checked for concurrent changes, and applied as atomic daily transactions across eight related tables. Exact before snapshots and a resumable receipt preserved rollback. Separate recovery tests check exact restore behavior and transaction safety.
A request count proves a recorded server request. It does not prove indexing, citation, referral traffic, ranking influence, or human demand.
Outcome
The derived crawler measurement layer is recovered for 71 verified dates. Missing evidence remains visible as unavailable. Current regression checks cover scanner exclusion, successful content retrieval, whole-count arithmetic, IPv4 and IPv6 range handling, and spoof evidence.
The durable result is not one repaired dashboard. It is a clearer system boundary: raw evidence is protected, derived products are replaceable, and any future rebuild must be reviewable and reversible.
See More Evidence
More SEO proof from live websites, each one with the numbers attached.