Idle warehouses, queries running on the wrong compute, table scans that should have been pruned, and AI calls on a model three tiers larger than the job needs. Each one found, priced, and handed to you as the exact statement that fixes it.
No Snowflake account needed — every new sign-in starts in a sample workspace with the full interface and realistic data.
| Warehouse | Idle | Verdict |
|---|---|---|
DEV_SANDBOX_WH X-LARGE | 31.4% | Downsize 184.2 credits/mo |
BI_REPORTING_WH MEDIUM | 12.8% | Scale-out review |
LEGACY_REPORTS_WH LARGE | 44.1% | Downsize 96.7 credits/mo |
ETL_INGEST_WH LARGE | 8.2% | Optimal |
Sample workspace · savings are estimates, each shown with its confidence
Nothing here is a score or a grade. Every finding names the measurement it came from, what it would cost to leave alone, and the statement that changes it.
Idle credits are measured as metered minus attributed — the gap between what you were billed and what a query was actually doing. Four levers come out of it, evaluated first-match-wins so you never get two fixes that invalidate each other.
Every query lands in one of four buckets — by tag first, then by text, statement type, and shape. Repeated patterns are then checked against the warehouse they ought to run on, and priced both ways.
Six detectors over the query text and its measured behaviour. Severity comes from the measurement, not the regex — a function on a filter column is only a finding when the pruning actually failed.
Cortex spend per model against its tier's credit rate, with an alert where a large model is doing work a cheaper tier does as well — and the specific model to downgrade to.
A recommendation you can't check is just an opinion with a number attached. Every one of these is checkable.
The app reads ACCOUNT_USAGE and INFORMATION_SCHEMA. Every change it suggests is shown as the exact statement, with its reverse next to it, and nothing runs until you choose to run it.
Idle is metered minus attributed, and both halves are shown. Savings are priced from the warehouse-size credit ratio and labelled with the confidence that model deserves — never rounded up into a headline.
Under 20 queries in the window, a warehouse reports Insufficient data rather than a guess. Where attribution is missing, idle is marked estimated instead of measured.
ACCOUNT_USAGE lags up to three hours and the interface says so on every screen that depends on it, with a live INFORMATION_SCHEMA tail where the account supports one.
Sign in and you land in a sample workspace — eight warehouses, every verdict, every finding. Connect your own when you're convinced.