A Snowflake data platform spending $262,887 a year had up to $101,129 in annual savings available. We found it in three weeks with read-only access—without changing the data model.

Names and object counts have been generalised to protect the client. All savings figures are exact and based on billed dollars.

Results

The committed savings required six configuration changes and one bug report. No application changes or change window were needed.

Key findings

Personal sandbox apps left compute running — $14,429/year

Six personal dashboard apps stayed active for up to three days after use. Reducing their auto-suspend setting to one hour would save an estimated $7,215/year with a 30–60 second cold start when reopened. This was suitable for sandbox work, not customer-facing services.

Staging data had 90-day retention — $5,822/year

A 35 TB staging database held almost 10,000 tables with 90-day Time Travel. Much of the data had not changed for more than 90 days. Once the owner confirms it can be reloaded from source, reducing retention is a simple configuration change.

A retention regression was caught early

Most databases used one-day retention, but newer databases had inherited 90 days from a changed provisioning template. The immediate cost was small; fixing it early prevented a larger storage bill later.

One warehouse was oversized for its workload — $14,176/year

A warehouse running many small, serial queries had an estimated 70% optimisation opportunity. It requires a data science team to change its code, so we classified it as probable rather than committed.

Assistant licences could be better targeted — $7,599/year

The top 50 users generated about 75% of usage. Limiting licences to that group was a commercial decision, not an engineering change.

The renewal was the largest opportunity — $26,289–52,577/year

The contract did not match how the platform was being used. Unlike configuration savings, this opportunity had a deadline: the renewal window.

What did not need work

A useful review also rules out false alarms. Warehouse auto-suspend, container suspension, clustering, pipes, and replication were already well configured. Developer databases and millions of dropped tables looked large in count, but were not financially material. One apparent spend spike was genuine growth, not waste.

What the review taught us

Our initial storage analysis was wrong. A table-level view under-reported account storage by more than 50 TB, so we reconciled the data against database-level history and the invoice. That work revealed the staging-retention finding above—worth $5,822/year—and became a mandatory check in our process.

How we priced the savings

We priced findings from billed dollars, not credits multiplied by assumed rates. Cloud-services costs were excluded where they would double-count the same fix. Each recommendation was assigned a confidence tier so the client could separate immediate action from opportunities that needed a business decision.

---

The dollar amount will differ in every account. The question is whether you know which few changes would make the biggest difference in yours.