Databricks Cost Insights & Optimization Actions
Insights brings cost anomalies, waste, and configuration opportunities into one review flow. Start with the evidence of a finding, then decide whether to investigate, defer, dismiss, or apply a change.
The Insight Type Catalog lists every finding, when it appears, and how to respond.
Active
Section titled “Active”Open findings for review.
- Views: All, Storage (tables and queries), and Resources (clusters, jobs, pipelines, warehouses, Apps, AI services, and Lakebase).
- Filters: search by title, category, severity, and type. Storage adds Catalog and Schema filters. Resources adds a resource-type filter. The global workspace filter also applies.
Snoozed findings stay in this list and in its counts.
History
Section titled “History”Findings of all statuses for the selected period. Use the status filter to show User Resolved, Action Taken, Auto-Resolved, or Dismissed findings. A resolved finding does not prove that the estimated saving appeared on an invoice.
Insight categories
Section titled “Insight categories”| Category | Purpose |
|---|---|
| Waste | Usage that may be avoidable, such as idle compute, repeated failures, or excessive scanning. |
| Anomaly | Departures from a baseline: cost, duration, failure rates, budgets, and attribution. |
| Optimization | Configuration or workload changes: worker sizing, runtime updates, schedules, and storage lifecycle. |
Severity shows urgency within one finding type. Confidence describes the strength of the evidence. Neither means that a proposed change is safe without review. An informational sizing finding can recommend keeping the current configuration.
Insight detail
Section titled “Insight detail”The detail page of a finding shows its summary, resource, evidence, savings estimate when available, action plan, and timeline. Evidence can include utilization, baseline and recent costs, sample counts, and recommended worker counts.
Check that the evidence describes the right resource and a relevant period. Each finding type uses its own observation window. The time filter of the page does not change it.
Choosing a response
Section titled “Choosing a response”| Response | What it does | Where |
|---|---|---|
| Mark Resolved / Resolve | Closes the finding after you verified a fix. | Mark Resolved on the detail page, or Resolve for selected rows in Active |
| Dismiss | Closes the finding without marking it fixed. It can return if the condition persists. | Select rows in Active |
| Snooze | Defers the finding for 1 hour to 30 days. Its evidence can still update. LakeSentry does not execute its plan while it is snoozed. | Select rows in Active |
| Suppress | Hides the open finding and prevents a new finding for the same condition while it stays open. Suppression has no undo. | Detail page |
| Approve / Reject | Accepts or rejects the action plan of the finding. | Active table, detail page, or selected rows |
Use Snooze for a temporary exception. Use Suppress for a finding you accept as is.
Only admins and owners can change findings, approve or reject plans, and use Recalculate. Recalculate queues a new detection run; the list updates when the run finishes. Bulk Dismiss, Resolve, and Snooze skip findings that are already snoozed.
How findings update and close
Section titled “How findings update and close”A detection run updates an open finding instead of creating a duplicate. A dismissed or resolved finding can return if the condition returns or persists.
LakeSentry closes findings automatically in these cases:
- Not detected again. The last detection is more than 7 days old (anomalies), 3 days old (waste), or 14 days old (optimization). This does not prove that the resource stopped costing money. Snoozed findings are excluded.
- The condition ended. The configuration changed, the resource disappeared, or the finding no longer qualifies.
Changes in Databricks appear after the next data extraction and processing run. History follows the retention policy of your deployment.
Savings estimates
Section titled “Savings estimates”Estimates help you set priorities. They derive from recent observed cost and assumptions specific to the finding type. They are not an invoice guarantee.
- The cost display mode (Total, Databricks Only, Cloud Only, Breakdown) changes the amount shown for a finding.
- Several findings can describe the same resource. Do not add their estimates together.
- A blank estimate can be correct. Billing data can be missing, or the finding gives advice without an amount. For example, a Spot Candidate has no estimate, because the bill of a cluster does not separate worker cost from driver cost. Some sizing findings quote savings only when measurements show that the cluster stayed at its minimum size.
Common questions
Section titled “Common questions”Why is a finding missing? Check the filters, data freshness, and the requirements in the catalog. A finding needs enough history. Missing telemetry is not zero usage. A suppressed finding is hidden.
Why did approval do nothing? Check the status of the plan on the detail page. LakeSentry executes only two plan types; see Action Plans. The others are proposals that you apply in Databricks. A supported action can also wait for the next execution cycle, or fail a safety check if the resource changed.