Fifteen minutes a month on the warehouse bill

Every month, someone pulls the top ten queries by spend and we look at them together. Fifteen minutes.

It has never found nothing. A sample of what it has found:

A dashboard nobody had opened since March, refreshing every fifteen minutes against raw facts. Roughly $600 a month for zero viewers.

A select * behind a BI tool on a 74-column table, of which four were used. Eighteen times the necessary I/O, 96 times a day.

A scheduled job with an hourly schedule whose consumer was a weekly report. Nobody had revisited the cadence since it was set in a hurry two years earlier.

A backfill someone had left running in a loop over a date range that had already completed.

A query whose partition filter was wrapped in a cast, defeating pruning entirely and scanning the full table history for one day of data.

That last one is the most instructive because the query looked correct:

-- Reads every partition, ever.
where cast(event_date as string) like '2026-03-14%'

-- Reads one.
where event_date = date '2026-03-14'

The general lesson is not about any of these specifically. It is that cost regressions are invisible in code review and obvious in a bill. Nobody reviewing that cast would flag it; the query is correct and readable. Only the spend report shows it.

Fifteen minutes, monthly, with two people. It is the highest return per unit of meeting time of anything we do.