Good questions.

Almost nothing. Point Datainspec at a version of your data that you trust. It profiles every column on its own: it works out which column is the key, which are categories, measures, ids, dates or free text, learns their normal ranges, and picks the checks that fit each one. There is no rules file to write before your first scan. If it gets a column wrong, you correct it in one click, and the correction is recorded.

CSV files (including a folder of CSV parts), Parquet, DuckDB and SQLite out of the box. Other SQL databases, such as PostgreSQL, connect through their own driver, which we set up with you. More sources are on the way.

On your own laptop, workstation or server, on Windows or Linux: as a desktop app, from the command line, or in a Docker container, so it fits into a pipeline or onto a server.

Windows 11 (64-bit) or Linux. The app brings its own engine, so there is nothing else to install; on Linux, the desktop build is made for your pilot, and the command line and Docker image are ready as they are. Memory depends on the size of your data: before a large scan it estimates what it needs and tells you whether it fits.

Yes. A tray icon shows each gate's state, and Windows notifies you when a gate blocks or a scan finishes. In a pipeline, the exit code tells your orchestrator.

Before a large scan it estimates the memory it needs and tells you whether it fits. On France’s company register, 30 million rows a side, a full September-against-October scan ran in about 21 minutes on a laptop.

From the version you approve. Each column’s thresholds come from its own data: its range, how often it is empty, how its values are spread. A price column and a status column are not judged with the same yardstick.

Yes. Every threshold is visible per column. Override any of them, say why, and revert whenever you like. Overrides are kept in the audit log and apply from the next scan.

Yes: ranges such as quantity ≥ 1, and lists of allowed values, written in a form or in YAML. It suggests rules from what the approved data showed, checks a new rule against that data before you save it, and can import rules from dbt or Great Expectations.

A health check of the version itself: duplicate keys, empty or constant columns, numbers stored as text, and placeholder codes such as 9999.9. Warnings are shown before you approve; a critical one has to be acknowledged. Every approved version is kept with who approved it and when, you can compare two versions, and a rollback adds a new version instead of erasing anything.

No. Real data changes for good reasons too. You mark each finding as a real problem, an expected change or noise, or snooze it. Your decision is recorded with your name and shown again when the same finding comes back.

No. Datainspec runs where your data is and sends nothing to us: no data, no profile, no finding, and no telemetry.

No. It only reads your sources, with read-only access enforced in layers and tested against hostile queries. Its own records and reports are stored separately.

The audit log keeps every approval, block, rollback, role correction and threshold override, with the person, the time and the reason. Filter it and export it as CSV. Each approved version also has its own approval certificate.

Yes. A share-safe export contains no values from the data: counts, rates and statistics only, with ids and values withheld.

Yes. Run it from the app, or from your own CI. Exit code 0 means clear, 1 warning, 2 blocked, 3 the scan failed. Datainspec gives the verdict; your pipeline decides what to do with it.

The checks are statistical and rule-based, and each finding shows its evidence. Explanations are written from templates, or by a local model on your machine if you install one.

A few weeks on one recurring dataset, on your own infrastructure, with success criteria agreed in writing before it starts. At the end we review together what it caught, what was noise, and whether you keep it.