Security and data
Know what leaves the environment—and what does not.
The collector is customer-controlled. The verified assessment ZIP contains workload metrics and configuration evidence, not the connection credentials used during collection.
Measured collector impact
What running the collector did to a test database
We measured the collector on Amazon RDS for SQL Server (db.r5.8xlarge, SQL Server 2022 Standard Edition, io1 storage) under a steady test workload (HammerDB, four virtual users). Each of six repetitions compared about one hour with and without the collector, in varied order, in September 2026.
How we tested
| Database | Amazon RDS for SQL Server db.r5.8xlarge (32 vCPU, 256 GiB memory), SQL Server 2022 Standard Edition (16.00.4095.4), io1 storage: 2,000 GiB with 5,000 provisioned IOPS. |
| Workload | HammerDB, a standard database load-testing tool, running an order-processing workload with four virtual users, driven from a separate test machine in AWS. |
| Each run | 5-minute ramp-up, about 60 minutes measured, then a 5-minute cooldown. |
| Three conditions | A: workload only. B: workload plus SQL Server measurement tools (Query Store and Extended Events), collector off. C: the same as B with the collector running. |
| Collector effect | C compared with B, so the measurement tools' own overhead is not counted against the collector. |
| Iterations | 6 repetitions of all three conditions: 18 measured runs, about 18 hours of measured workload (about 21 hours including ramp-up and cooldown). |
| Order | The condition order changed every repetition (ABC, ACB, BAC, BCA, CAB, CBA), so slow drift in the workload over time does not favor one condition. |
| Results | The median of the six paired C-minus-B differences, with the lowest and highest repetition shown in the full report. |
| When | September 28 to 29, 2026. An earlier three-repetition round in a fixed order was inconclusive and was replaced by this counterbalanced round. |
What we measured
- CPU: no increase measured.
- Memory: no increase in memory pressure.
- Connections: effectively unchanged.
- Space on your server: about 0.5 MB of collector working space, removed when collection ends.
- Disk activity: a small, expected increase in reads (the collector reads SQL Server's own performance views); no increase in writes.
- Application throughput: in five of six repetitions, the test application completed fewer transactions while the collector ran (median about 6.6% fewer; one repetition completed more). The cause is not yet determined. We are investigating and will publish updated results here.
What this means for you: these results come from one test setup, and your workload may behave differently. The collector never reads or changes application data. If your application is sensitive to small slowdowns, run the first collection on a non-production copy or start it during a quieter period, and watch your usual monitoring.
How your account is protected
Accounts, access, and records
- You sign in with a one-time link sent to your work email, so you never create or enter a Savio password.
- Each collector you download carries a personal token linked to your account. An upload is accepted only when you are signed in to the account that owns that token, so a copied collector cannot be used by someone else.
- Your reports and account details are visible only to your account. Only authorized Savio staff may open assessment results, and only to provide support.
- Collector tokens are stored only as one-way fingerprints, never in readable form. Failed sign-in checks, collector downloads, rejected uploads, and registration are recorded.
- Savio does not make AWS, database, or configuration changes. You decide whether and when to act on a result.
Hosting: encrypted connections (HTTPS) and storage encryption are provided by Savio's hosting and database providers. Questions about security: [email protected].
Review the collection method
See how evidence is converted into the assessment result.
Review Methodology →