How it works
Five steps, and one rule underneath all of them: never show a number whose derivation isn't on the same screen.
- 01
Read
Two sources. ACCOUNT_USAGE gives metered credits per warehouse-hour, per-query attribution, and query history — authoritative, and up to three hours behind. INFORMATION_SCHEMA gives a live tail for the last 24 hours where the account exposes it. The app tags its own queries so its cost is visible and excluded from the workload it is analysing.
- 02
Measure idle
Idle is metered credits minus the credits attribution could tie to a running query. Both halves are shown, and they always reconcile to the total — attribution that falls outside its metered day, from queries running across midnight, is clamped rather than dropped. Where attribution is unavailable, idle is estimated from execution time per warehouse-hour and labelled estimated, not measured.
- 03
Decide
Four thresholds, evaluated first-match-wins: under 20 queries is insufficient data; sustained queueing is a scale-out question; high idle with a long auto-suspend is a suspend question; fast, small queries on a warehouse that has a smaller size available is a downsize. The order is deliberate — fixing auto-suspend changes the numbers the downsize decision would have been made on, so it is never recommended at the same time.
- 04
Price
A warehouse size step doubles the node count and so the credit rate. Savings are the credit-rate ratio applied to the observed window and scaled to a month. That model assumes runtime scales inversely with nodes, which is optimistic for most real queries — so every figure derived from it is labelled MEDIUM confidence, on screen, next to the number.
- 05
Hand it over
Each recommendation renders as the statement that applies it and the statement that reverses it. Applying is something you do, not something that happens. Downsizing moves one size step at a time on purpose: reaching the right size over a few cycles with evidence at each step beats one large jump that nobody can defend.
What it will not do
- Guess on thin data. Under 20 queries in the window, a warehouse reports insufficient data.
- Round an estimate up into a headline. Confidence travels with the number everywhere it appears.
- Change anything without being asked, or hide a change that costs more than it saves.