top of page

BigQuery row-level security and data masking now GA on Iceberg tables

4 hours ago
7 min read
BigQuery row-level security and data masking now GA on Iceberg tables, Data Studios original cover

Google Cloud has made row-level security and column-level security with dynamic data masking generally available for Apache Iceberg tables managed by the Lakehouse runtime catalog. The change is listed in the BigQuery release notes under November 18, 2026, and it means a data team can now protect open-format Iceberg tables with the same kind of row filters and masked columns that BigQuery already offered for its native tables, with production support behind it.


The release note adds two details that decide how far the protection reaches. Policies are managed with GoogleSQL DDL in BigQuery, and on catalogs that use credential vending, the Apache Iceberg REST catalog endpoint hands out storage credentials only to principals that have unmasked access to every protected column and unfiltered access to every row. A user who is filtered or masked in BigQuery therefore cannot sidestep the policy by pointing Apache Spark, Trino or Apache Flink at the same table and reading the Parquet files directly.


The Lakehouse runtime catalog is the product Google previously called BigLake metastore. The documentation says it was renamed on April 20, 2026, and that IAM role and permission names still use the BigLake prefix, which explains why the roles in this setup are called roles/biglake.admin and roles/biglake.viewer.


··········


WHAT GOOGLE MADE GENERALLY AVAILABLE FOR LAKEHOUSE ICEBERG TABLES.


The feature covers three controls on Iceberg tables in the Lakehouse runtime catalog. Row access policies filter which rows each principal can see. Column data policies of the masking type return a transformed value instead of the original. Column data policies of the raw-access type let named principals read the original value of a protected column. Google's launch-stage definitions describe general availability as ready for production use, so the feature now carries the support commitments of a GA product rather than the more limited terms of a preview.


The documentation page for the feature, titled "Row-level and column-level security" and last updated on November 18, 2026, sets out what is covered and what is not. Fine-grained access control applies to Iceberg tables managed by the Lakehouse runtime catalog. It is not supported on federated catalogs or on external tables accessed through cross-cloud connections.


........


Item

What the documentation states

Status

Generally available (GA), release note dated November 18, 2026

Tables covered

Apache Iceberg tables in the Lakehouse runtime catalog

Not covered

Federated catalogs, cross-cloud external tables

How policies are managed

GoogleSQL DDL only

Not supported for bindings

Console, tables.patch, bq CLI, Terraform

Policy tags

Data Catalog policy tags and governance tags not supported

Masking expressions shown

SHA256 and ALWAYS_NULL

Open-source engines

Spark, Trino, Flink through the Iceberg REST catalog endpoint


........


The DDL-only restriction has practical weight for teams that manage BigQuery governance as infrastructure code. The documentation says the console, the tables.patch API method, the bq command-line tool and Terraform are not supported for these column bindings, so policy changes on Iceberg tables have to be scripted as SQL statements and kept under version control in that form. Teams that already store their BigQuery SQL in BigQuery Studio Git repositories can keep these policy scripts next to the rest of their transformation code.


The second restriction concerns teams that classify columns with policy tags. On native BigQuery tables, column-level security is often built on Data Catalog taxonomies. For Iceberg tables in the Lakehouse runtime catalog, the documentation says policy tags and data governance tags are not supported, and the column policies are attached directly to each column as data policies.


··········


HOW ROW ACCESS POLICIES AND DATA POLICIES WORK ON ICEBERG TABLES.


The setup follows a fixed order in the documentation. Billing must be enabled on the project, the BigLake API and the BigQuery Data Policy API must be enabled, the Lakehouse runtime catalog must be configured with the Apache Iceberg REST catalog endpoint in credential-vending mode, and an Iceberg table must exist. Data policies and row access policies are then created, column policies are attached, and grants are issued. Tables are always addressed with a four-part identifier made of project, catalog, namespace and table name.


The statements below follow the syntax shown in Google's documentation, with placeholder names for an illustrative orders table.


CREATE DATA_POLICY IF NOT EXISTS `my-project.region-us.email_hash`
OPTIONS (
  data_policy_type = "DATA_MASKING_POLICY",
  masking_expression = "SHA256"
);

CREATE DATA_POLICY IF NOT EXISTS `my-project.region-us.email_raw`
OPTIONS (
  data_policy_type = "RAW_DATA_ACCESS_POLICY"
);

GRANT FINE_GRAINED_READ ON DATA_POLICY `my-project.region-us.email_hash`
TO "principalSet://goog/group/analysts@example.com";

GRANT FINE_GRAINED_READ ON DATA_POLICY `my-project.region-us.email_raw`
TO "principalSet://goog/group/finance-leads@example.com";

ALTER TABLE `my-project.sales_catalog.retail.orders`
ALTER COLUMN customer_email SET OPTIONS (
  data_policies += ["{'name':'my-project.region-us.email_hash'}",
                    "{'name':'my-project.region-us.email_raw'}"]
);

CREATE OR REPLACE ROW ACCESS POLICY us_rows
ON `my-project.sales_catalog.retail.orders`
GRANT TO ("group:us-analysts@example.com")
FILTER USING (region = "US");

One detail in the documentation catches people out. Grants on data policies use IAM v2 principal identifiers such as principalSet://goog/group/, while the GRANT TO clause of a row access policy uses the standard IAM v1 identifiers such as group: and user:. A policy script for one table therefore has to use both formats, each in its own statement type.


Access then resolves in a predictable way. A principal with FINE_GRAINED_READ on the masking policy sees the masked value; a principal with FINE_GRAINED_READ on the raw policy sees the original. A principal with no grant on any data policy attached to a protected column gets an access-denied error when the query touches that column, although the unprotected columns stay queryable when the query selects every column except the protected one. For rows, a principal that appears in no row access policy sees zero rows, and a principal listed in several policies sees the union of the rows those policies allow. The documentation also states that subqueries are not supported in the FILTER USING expression.


··········


ILLUSTRATIVE EXAMPLE: WHO SEES WHAT IN A PROTECTED ICEBERG TABLE.


The example below is illustrative. It uses a six-row orders table with a region column, a customer_email column and an amount column worth 4,300 in total. Three row access policies exist: us_rows for the us-analysts group, eu_rows for the eu-analysts group, and full_rows with FILTER USING (TRUE) for the finance-leads group and an ETL service account. The customer_email column carries the SHA256 masking policy, granted to analysts, and the raw-access policy, granted to finance leads and the ETL service account. The results were computed in Python by applying the documented rules.


........


Principal

Groups

Rows visible

Sum of amount

customer_email

Ana

us-analysts

3 of 6

2,300

SHA-256 hash

Ben

eu-analysts

2 of 6

1,350

SHA-256 hash

Dev

us-analysts and eu-analysts

5 of 6

3,650

SHA-256 hash

Cara

finance-leads

6 of 6

4,300

Original value

Eve

none

0 of 6

0

Access denied


........


Dev's result shows the union rule at work: membership in both analyst groups yields the three US rows plus the two EU rows, while the single APAC row stays hidden because no policy granted to Dev covers it. Eve's result shows the default-deny behavior, which is the main reason to add a full-access policy for the people and service accounts that need the whole table before a first row policy goes live.


The same matrix decides what happens outside BigQuery. If Ana runs a Spark job against the table through the Iceberg REST catalog endpoint in credential-vending mode, the endpoint returns 403 Forbidden, because she is both filtered and masked. The ETL service account holds raw access on the protected column and a FILTER USING (TRUE) row policy, so it receives reduced-scope, short-lived storage credentials and can read and write the files. The hash also keeps joins possible for analysts: the same email produces the same SHA-256 value, so masked keys can still match across tables protected with the same expression.


··········


LAKEHOUSE ICEBERG SECURITY LIMITS AND BIGQUERY EDITION REQUIREMENTS.


The most important limit concerns catalogs that do not use credential vending. The documentation says that in that configuration BigQuery enforces the policies only on queries that run in BigQuery, and that with end-user credentials mode direct reads from Cloud Storage are left unprotected. For a table holding personal or regulated data, credential-vending mode is the configuration that makes the policies apply to every engine that goes through the catalog.


The required roles are listed precisely. Policy administrators need roles/bigquerydatapolicy.admin together with BigQuery and BigLake administrative roles such as roles/bigquery.admin, roles/bigquery.dataOwner or roles/biglake.admin, and readers need roles/biglake.viewer and roles/bigquery.jobUser to run queries. The system-managed role roles/biglake.filteredDataViewer is assigned through the grantee lists of row access policies, and the documentation warns against granting it directly.


The Lakehouse page itself does not state an edition requirement or a separate price. BigQuery's editions documentation lists fine-grained security controls, including column-level access control, row-level security and dynamic data masking, under Enterprise and Enterprise Plus, and states that Standard edition has no access to them. Teams running Standard edition reservations should confirm which reservation will run the queries before planning a rollout, and the trade-offs between capacity and on-demand billing are covered in our explainer on BigQuery on-demand and editions pricing. The column masking documentation adds an indirect cost effect: columns masked with nullify or default-value rules are not scanned, which can lower bytes processed for masked users.


Other engines approach the same problem in different ways. Snowflake, for example, opened Iceberg tables to outside readers through its own catalog work, described in our report on Snowflake Open Data Sharing for Iceberg tables. Google's approach puts the enforcement point at credential issuance, so the decision about who may read raw files is taken by the same policies that govern BigQuery queries.


··········


BIGQUERY ICEBERG ROW AND COLUMN SECURITY: WHAT DEPENDS ON YOUR CATALOG SETUP.


Google confirms that row access policies and masking and raw-access data policies are generally available on Iceberg tables in the Lakehouse runtime catalog, that they are managed only with GoogleSQL DDL, and that the Iceberg REST catalog endpoint withholds storage credentials from anyone who lacks full access to a protected table. The masking expressions shown for these tables are SHA256 and ALWAYS_NULL, and policy tags are not part of this model.


How much protection a team actually gets depends on its configuration. With credential vending enabled, the same rules apply to BigQuery queries and to Spark, Trino or Flink jobs that go through the catalog. Without it, only BigQuery queries are filtered and masked. The edition behind the reservation, the IAM format used in each statement, and a full-access row policy for pipelines and administrators are the three items to settle before attaching the first policy to a production table.


Google's documentation for this release does not list a separate charge for the feature or a per-region availability table, so teams with data residency requirements should confirm the catalog location and the data policy location, which the examples express as a region-qualified name, in their own project before migrating sensitive Iceberg tables.


··········


FOLLOW US FOR MORE.


·····


DATA STUDIOS


·····


datastudios.org


bottom of page