Zero-trust started as a network idea: verify every request. Applied to data platforms it means never assuming — not that upstream data is clean because it was clean yesterday, and not that access granted eighteen months ago is still appropriate today.
The manual alternatives are the ones that fail. Broken pipelines silently feeding wrong numbers to dashboards cost real money and — worse — credibility, and the spreadsheet of who-can-see-what stopped reflecting reality two reorgs ago. The fix for both is the same: move the checks into the pipeline, as code.
Quality checks that block, not report
The zero-trust posture for quality is validation on every state change, with failures that stop the pipeline rather than emailing a report nobody reads. Great Expectations makes expectations executable — column X is never null, values of Y fall in this range, row counts stay within tolerance — and runs them as gates between pipeline stages, so bad data is quarantined before it propagates.
Declared checks catch what you anticipated. Observability platforms like Monte Carlo cover the failures you didn't: they learn each table's normal freshness, volume, and distribution, then alert on anomaly — the feed that silently halved, the schema change nobody announced. Detecting an issue before a stakeholder does is the difference between a fix and an apology.
Access control as code, via dbt
Hand-run GRANT statements drift. Since dbt already knows every model it builds, it is the natural place to declare access: grants in model configs mean every dbt run reapplies the intended permissions — rebuilt tables never come up unprotected, and revoking access is a pull request with an audit trail, not a ticket.
Structure the grants around roles, not people: analyst, finance, pii-reader. Tag models containing sensitive columns and let policies key off the tags, so a new PII table is protected by convention rather than by someone remembering.
Governance people can live with
Zero-trust fails when it becomes zero-productivity. The goal is defaults that are safe and paths that are fast: requesting access is a small PR, quality failures point at the exact expectation that broke, and the lineage graph shows who is affected before you ship a breaking change. Governance that helps engineers ship gets adopted; governance that only says no gets bypassed.