Data governance is no longer a back-office concern. Analysts, pipelines, and AI applications all read from the same lakehouse, and each new consumer widens the set of people who can change, expose, or misuse a dataset. Unity Catalog is the control plane Databricks provides for this, and Buoyant Data helps teams configure it so that access, lineage, and audit answer the questions that data quality and compliance reviews actually ask.
Why governance and data quality are the same problem
Bad data is usually a governance failure that has not been noticed yet. In the RSM Middle Market AI Survey 2025: U.S. and Canada, which surveyed 966 respondents at middle-market firms in the United States and Canada, 41% of respondents who experienced AI implementation issues cited data quality as their top problem. When a model or dashboard is wrong, the first questions are: where did this data come from, who changed it, and who was allowed to read it? Governance is what makes those questions answerable.
Grants by group
Unity Catalog grants privileges on catalogs, schemas, tables, volumes, and functions to groups rather than to individual users. A well-run deployment usually has:
- Groups aligned to responsibilities, such as data producers, analysts, auditors, and service principals for pipelines, synchronized from your identity provider so that joiners and leavers are handled in one place.
- Least-privilege defaults, with write access limited to the pipelines that own a table and read access granted per consumer group.
- Row filters and column masks for personal or regulated fields, so that one table can serve several audiences without copies.
- Reviewable grant state. Permissions are defined in code or exported on a schedule, so that a compliance reviewer can see who could read a sensitive table on any given day, not just who was granted access last quarter.
Lineage
Table and column lineage recorded by Unity Catalog show how data moves from raw ingestion through transformations into reports and models. Lineage is what turns a data quality incident into a bounded investigation:
- Trace a bad metric back to the upstream tables and jobs that produced it.
- Find every downstream dashboard, notebook, or model that depends on a column before it is renamed, retyped, or dropped.
- Show a regulator or auditor the provenance of a reported figure without reconstructing it by hand.
Lineage is only as good as the workloads that run through Unity Catalog. We check that pipelines write through governed tables and that external sources are registered, so that the graph is complete rather than partial.
Audit trails
Databricks publishes access and activity records to the system.access
schema, including audit events and table lineage, which can be queried like any
other table. We set this up so that:
- Access to sensitive catalogs is queryable by user, group, and time window.
- Grant and revoke changes are logged and retained for the period your compliance policy requires.
- Alerts or scheduled reviews flag unusual access patterns, such as a service principal reading a table it has never read before.
The audit data is itself governed. Only the compliance group should be able to read it, and retention is set explicitly rather than left at defaults.
How we work
A governance engagement is scoped, not open-ended. A typical sequence is:
- Assessment: inventory of catalogs, schemas, grants, and the groups that own them, with the sensitive data and the audiences that need it identified.
- Group and grant model: a proposed structure of groups and privileges, reviewed with your data owners before anything changes.
- Implementation: grants, row filters, and column masks applied in non-production first, then promoted with a rollback plan.
- Audit and lineage handoff: queries, dashboards, and a review cadence your team can run without us, with documentation for the next person who asks who can see what.
Governance should answer the audit question before the audit happens.
Ready to tighten access and lineage?
Contact us