Confidence is evidence strength, not certainty
High, medium, and low confidence summarize how strongly the available evidence supports the finding. Confidence is not a guarantee that a change is safe for your workload, and it does not replace review of the proposed SQL, risk notes, or maintenance requirements.
Fix zones describe the required control
We can fix now
The recommendation has a concrete eligible change. Project permissions and all runtime safety gates still decide whether it can run automatically.
Suggested SQL
Review the SQL and have an authorized database operator run it through the normal change process.
Needs restart / review
The change has operational impact, such as a database server configuration or restart requirement, and needs planned review.
Manual only
The evidence calls for investigation or work that dbtempo cannot safely express as an executable change.
Read the evidence and provenance
- Runtime evidence comes from observed query statistics, schema, indexes, database configuration, or execution plans.
- Code evidence comes from an indexed source location associated with the query.
- Mixed evidence means the recommendation combines observed runtime behavior with a source-code location.
When before and after EXPLAIN evidence is available, compare access paths, selected indexes, estimated rows, sorts, and temporary work. Missing after evidence does not invalidate the finding; it means the proposed plan was not available for comparison. Automated probes run the query only on PostgreSQL, and only after a project owner turns on query execution for the project.
Make and record the decision
- Apply queued fixes from the approval queue. When a project capability is set to "Ask me first", eligible fixes collect under the recommendations page's "Awaiting approval" tab. Apply executes the fix through the durable manual-apply path and shows the outcome on the card; the Audit tab keeps the full record.
- Save it for later. Later parks the recommendation without executing anything — it stays applicable from its detail page. Use "Mark as applied" there if you ran the change yourself.
- Dismiss when it does not apply. Dismiss removes the finding from the open queue. Restore it later when the workload or your decision changes.
- Verify after a change. Re-run the representative workload or scheduled analysis and compare the new plan and performance signals with the original evidence.