Snowflake cost-fit assessment
In about a week I go through every query in your account, score each one by how much it actually spilled to disk, and reconcile the totals to your invoice. You get a clear picture of what could run on a single DuckDB machine, what should stay on Snowflake, and roughly what you'd save. It's a fixed fee, and the scoring is open-source SQL you can read for yourself.
About 15 minutes, read-only, runs on your laptop. Nothing leaves your account.
Example breakdown · acme-prod, 90 days
38%of the bill is query compute that fits on one machine
Most Snowflake cost tools stop at showing you where the money goes. This one works out what could actually run somewhere cheaper.
Tested on Snowset
Snowset is a public trace of real Snowflake production queries. I ran ducklens over a 5.8 million query sample of it.
Who this is for
The job is to prove what's movable, in writing, before anyone touches production.
How it works
We give you the exact SQL to run against ACCOUNT_USAGE. You run it yourself, so
we never see your credentials and nothing touches production.
The open-source SQL scores each query on whether it would really fit on one machine, based on how much it spilled to disk rather than how much it scanned. The totals reconcile to your invoice.
A move, split, or keep call for each warehouse, a saving range, and what a migration would cost. Sometimes the answer is that it isn't worth it, and I'll tell you that too.
For the engineer who gets paged if it breaks
Every call the tool makes comes from one SQL file you can read. The thresholds are all named settings you can change and re-run if you disagree with one.
The headline number is anchored to METERING_DAILY_HISTORY, so it has to match
your Snowsight bill. If it doesn't, something is off, and you'll see where.
Write-heavy work, high-concurrency serving, regulated data, and Snowflake-specific SQL all stay on Snowflake. The report lists each one with the dollars attached and the reason it's staying.
Before a single dashboard is repointed, a parallel-run check compares the old and new results on a frozen snapshot. If one row is off, it fails.
Your ACCOUNT_USAGE export and your credentials never leave your machine, so
there's nothing for a third party to put through a security review.
Snowflake's renewal pricing has been going up 15 to 22% a year for existing customers. Even if you decide to stay, having a scored exit plan in hand helps in that conversation.
The offer
The free tool gives you a rough number on your own data. The paid assessment gives you the version reconciled to your invoice, a clear recommendation, and someone accountable for it.
Bill read-out
You send the export, or we run the scan together, and I walk you through your fit report live. Credited in full against the assessment.
Assessment
from $7,500
fixed fee · ~1 week
The decision, on your real 90-day history, reconciled to your invoice.
If it doesn't find at least its own fee in savings, or it tells you honestly to stay, you don't pay. You keep the report either way.
Pilot recommended
$8–20k
~2–4 weeks · fixed fee
One real workload moved to DuckDB and DuckLake, checked against your data.
Migration
$20–50k
scoped from the pilot
We move what the assessment marked MOVE, and only that. You own the stack.
The honest part
Replacing Snowflake only makes sense for working sets under about 10 TB and under roughly 10 to 20 concurrent queries. If your workload is one of these, the assessment tells you to keep it, in writing.
Continuous MERGE and COPY, high-frequency ingestion. DuckDB is single-writer, so sustained writes hit a ceiling Snowflake doesn't have.
Dozens of concurrent dashboard users, multi-cluster scale-out. That's a serving workload, not single-machine batch.
Strict residency, multi-engine federation, Secure Data Shares. Above about 10 TB the honest answer is often Iceberg rather than us.
You migrate the part that genuinely fits, which is often less than half, and keep the rest. Knowing what to leave alone matters as much as knowing what to move.
How this is different
Optimizers tune your warehouses and take a cut of the savings for as long as you use them. This looks at a different question: whether most of the bill needs a warehouse in the first place.
There's no proxy sitting in your query path and no new service to depend on. It's a one-time assessment, and if you migrate, you own the result.
Adaptive Compute improves how many queries you get per dollar, but the price per credit is the same. This shows you which of those queries would cost almost nothing on hardware you own.
Cost dashboards show you spend broken down by warehouse and stop there. This runs each query against a real engine and gives you a call on it: move, split, or keep, with the dollars attached.
The deliverable
You get a one-page report you can hand to your finance team, reconciled to your invoice, with every threshold visible and the saving attached. It's the report itself, the same one you'd forward internally.
One page, reconciled to your invoice. Opens in full in your browser ->
The scoring is open-source SQL, MIT-licensed. Read every line before you run it.