Column-Level vs Row-Level Security in BigQuery

GCP Study Hub
October 26, 2026

BigQuery has two separate mechanisms for restricting access inside a single table. Column-level security controls which columns a user can query, and it is built on policy tags defined in Knowledge Catalog. Row-level security controls which rows a user gets back, and it is built on row access policies written in SQL. They solve different problems and are configured in different places, and that distinction is the main thing to hold onto for the Professional Cloud Architect exam.

Column-level security with policy tags

Column-level security starts in Knowledge Catalog, the data governance layer that sits above your various data services. Within Knowledge Catalog, you define sensitivity categories as policy tags. You might create a tag for PII (personally identifiable information), another for financial data, or whatever categories make sense for your organization. Policy tags are organized into taxonomies, and the BigQuery admin console also lets you create and manage policy tag taxonomies to classify and control access to sensitive data.

Once the policy tags exist, you attach them to specific columns in your BigQuery tables. A table might have four columns, with two of them tagged and two left untagged. Attaching a tag marks a column as requiring special permissions, and access control on that column applies automatically from then on. Users can no longer simply query a tagged column. They need explicit permission through the fine-grained reader role.

The fine-grained reader role is granted to users on the policy tags themselves. You could give someone access to all PII-tagged columns, or only to specific subcategories. This is what makes the approach scale well. Instead of managing permissions column by column across hundreds of tables, you manage them through policy tags, and a change to the permissions on a tag applies everywhere that tag is used.

Giving training jobs access to sensitive columns

The fine-grained reader role can be granted to service accounts as well as to people, which matters for machine learning workloads that need sensitive data. Granting a Gemini Enterprise Agent Platform (formerly Vertex AI) service account the appropriate fine-grained reader role lets training jobs read the tagged columns they need while your security controls stay in place for everyone else. Model training happens in a controlled environment, and the raw sensitive data never leaves your security perimeter or becomes visible to accounts that are not supposed to see it.

Row-level security with row access policies

Row-level security controls which rows different users can see within the same table, and it is configured differently. Row access policies in BigQuery are created with SQL statements. You define a policy that filters rows on some condition, then grant that policy to specific users or groups.

Consider a table of client data with an ID, a client name, revenue, and a segment column that classifies each customer as either small or enterprise. Several teams need access to the table, but they should not all see every client. A policy for the startup support team could be defined like this:

CREATE ROW ACCESS POLICY small_business_filter
ON mydataset.customer_data
GRANT TO ("group:startup-support@example.com")
FILTER USING (segment = "small");

When someone from the startup support group queries the table, they see only the rows where the segment is small, which in this example are TechStart and LocalBiz. They do not see MegaCorp, because it is an enterprise segment customer. A second policy, an enterprise filter, grants the enterprise sales group access to rows where the segment equals enterprise. That team sees the MegaCorp row and not the other two.

BigQuery applies these filters automatically behind the scenes. Users query the table normally and get back only the rows they are authorized to see based on the policies. They do not need to remember to add WHERE clauses, because the security is enforced transparently.

When an architect uses each one

Row-level security is useful for multi-tenant data, or whenever different teams should see only their relevant subset of the same table. Instead of creating separate tables or views for each team, you maintain one table and let row access policies control what each group can see. Column-level security addresses a different requirement, which is protecting specific sensitive fields, such as PII or financial data, across many tables with permissions managed centrally.

  • Column-level security restricts columns. It is defined with policy tags in Knowledge Catalog and enforced through the fine-grained reader role granted on those tags.
  • Row-level security restricts rows. It is defined with SQL row access policies in BigQuery, each pairing a filter condition with the users or groups it applies to.

For the Professional Cloud Architect exam, it helps to read a requirement in those terms. Hiding particular sensitive fields from most users points to policy tags and the fine-grained reader role. Letting each team see only its own records in a shared table points to row access policies.

Our Professional Cloud Architect course covers BigQuery column-level and row-level security alongside Knowledge Catalog and Sensitive Data Protection, with practice questions that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow