The Full Scan That Used to PruneWarehouse · medium · ~12 minAfter a 'code cleanup' PR, the hourly orders extract went from scanning ~55GB to scanning 4.1TB — the entire table — while returning identical results. On a bytes-billed engine this would be a 75x bill; on Snowflake it's 40x the runtime. The diff looks harmless: someone made the date filter 'more readable'.Find the pruning killer in the diff.Evidence

Diff & Scan Stats

Config

Checkpoints

Why did the readable version scan everything?

The general rule that fixes both diffs in the PR?

Results were identical, so tests passed. What guardrails catch pruning regressions? Interview-style.