Using row-level security with other BigQuery features

This document describes how to use row-level access security with other BigQuery features.

Before you read this document, familiarize yourself with row-level security by reading Introduction to BigQuery row-level security and Working with row-level security.

The `TRUE` filter

Row-level access policies can filter the result data that you see when running queries. To run non-query operations, such as DML, you need full access to all rows in the table. Full access is granted by using a row access policy with the filter expression set to TRUE. This row-level access policy is called the TRUE filter.

Any user can be granted TRUE filter access, including a service account.

Examples of non-query operations are:

Other BigQuery APIs, such as the BigQuery Storage Read API.
Some bq command-line tool commands, such as the bq head command.
Copying a table

Example

Create the `TRUE` filter

CREATE ROW ACCESS POLICY all_access ON project.dataset.table1
GRANT TO ("group:all-rows-access@example.com")
FILTER USING (TRUE);

Features that work with the `TRUE` filter

Copy jobs

To copy a table with one or more row-level access policies on it, you must first be granted TRUE filter access on the source table. All row-level access policies on the source table are also copied to the new destination table. If you copy a source table without row-level access policies onto a destination table that does have row-level access policies, then the row-level access policies are removed from the destination table, unless the --append_table flag is used or "writeDisposition": "WRITE_APPEND" is set.

Cross-region copies are allowed and all policies are copied. Subsequent queries might be broken after the copy is complete if the queries contain invalid table references in subquery policies.

Row-level access policies on a table must have unique names. A collision in row-level access policy names during the copy results in an invalid input error.

Required permissions to copy a table with a row-level access policy

To copy a table with one or more row-level access policies, you must have the following permissions, in addition to the permissions required to copy a table without a row-level access policy.

Permission	Resource
`bigquery.rowAccessPolicies.list`	The source table.
`bigquery.rowAccessPolicies.getIamPolicy`	The source table.
The `TRUE` filter	The source table.
`bigquery.rowAccessPolicies.create`	The destination table.
`bigquery.rowAccessPolicies.setIamPolicy`	The destination table.

Tabledata.list in BigQuery API

You need TRUE filter access in order to use the tabledata.list method in the BigQuery API on a table with row-level access policies.

DML

To execute a DML statement that updates a table that has row-level access policies, you need TRUE filter access for the table.

In particular, MERGE statements interact with row-level access policies as follows:

If a target table contains row-level access policies, then you need TRUE filter access to the target table.
If a source table contains row-level access policies, then the MERGE statement only acts on the rows that are visible to the user.

Table snapshots

Table snapshots support row-level security. The permissions that you need for the base table (source table) and the table snapshot (destination table) are described in Required permissions to copy a table with a row-level access policy.

Time travel is not supported for base tables that have one or more row-level security policies.

BigQuery table with JSON columns

Row-level access policies cannot be applied on JSON columns. To learn more about the limitations for row-level security, see Limitations.

BigQuery BI Engine and Looker Studio

BigQuery BI Engine does not accelerate queries that are run on tables with one or more row-level access policies; those queries are run as standard queries in BigQuery.

The data in a Looker Studio dashboard is filtered according to the underlying source table's row-level access policies.

Column-level security

Row-level security and column-level security, which includes both column-level access control and dynamic data masking, are fully compatible.

Key points are:

You can apply a row-level access policy to filter data in any column, even if you don't have access to the data in that column.
- Attempts to access these columns with the subquery row-level access policy results in an error indicating that access is denied. These columns aren't considered system-referenced columns.
- Attempts to access these columns with the non-subquery row-level access policy bypass column-level security.
If the column is restricted due to column-level security, and the column is named in the query's SELECT statement or subquery row-level access policies, you receive an error.
Column-level security also applies with a SELECT * query statement. The SELECT * is treated the same as a query which explicitly names a restricted column.

Example of row-level security and column-level security interacting

This example walks you through the steps for securing a table and then querying it.

The data

Suppose that you have the DataOwner role for a dataset named my_dataset which includes a table with three columns, named my_table. The table contains the data shown in the following table.

In this example, one user is Alice, whose email address is alice@example.com. A second user is Bob, Alice's colleague.

rank	fruit	color
1	apple	red
2	orange	orange
3	lemon	yellow
4	lime	green

The security

You want Alice to be able to see all the rows that have odd numbers in the rank column, but not even-numbered rows. You don't want Bob to see any rows, even or odd. You don't want anyone to see any data in the fruit column.

To restrict Alice from seeing the even-numbered rows, you create a row-level access policy which has a filter expression based on the data that appears in the rank column. To prevent Bob from seeing even or odd rows, you don't include him in the grantee list.
```
CREATE ROW ACCESS POLICY only_odd ON my_dataset.my_table GRANT
TO ('user:alice@example.com') FILTER USING (MOD(rank, 2) = 1);
```
To restrict all users from seeing data in the column named fruit, you create a column-level security policy tag that prohibits all users from accessing any of its data.

Finally, you also restrict access to the column named color in two ways: the column is governed both by a column-level security policy tag prohibiting all access by anyone, and is affected by a row-level access policy, which filters some of the row data in the color column.

This second row-level access policy only displays rows with the value green in the color column.

CREATE ROW ACCESS POLICY only_green ON my_dataset.my_table
GRANT TO ('user:alice@example.com') FILTER USING (color="green");

Bob's query

If Alice's coworker Bob tries to query data from my_dataset.my_table, he doesn't see any rows, because Bob isn't in the grantee list for any row-level access policy on the table.

Query	`my_dataset.my_table`			Comments
	`rank` (Some data is affected by the row access policy `only_odd`)	`fruit` (All data is secured by a CLS policy tag)	`color` (All data is secured by a CLS policy tag, and some data is affected by the row access policy `only_green`)
`SELECT rank FROM my_dataset.my_table`	(0) rows returned.			Bob is not on the row-level access policy's grantee list; therefore this query succeeds, but no row data is returned. A message is displayed to Bob that says his results may be filtered by the row access policy.

Alice's queries

When Alice runs queries to access data from my_dataset.my_table, her results depend on the query she runs and the security, as shown in the following table.

Query	`my_dataset.my_table`			Comments
	`rank` (Some data is affected by the row access policy `only_odd`)	`fruit` (All data is secured by a CLS policy tag)	`color` (All data is secured by a CLS policy tag, and some data is affected by the row access policy `only_green`)
`SELECT rank FROM my_dataset.my_table`	(2) odd-numbered rows are returned.			Alice is on the grantee list for the `only_odd` row-level access policy on data in the rank column. Therefore, Alice sees only the odd-numbered row data. Even-numbered rows are hidden by the row-level access policy named `only_odd`. A message is displayed to Alice that says her results may be filtered by the row access policy.
`SELECT fruit FROM my_dataset.my_table`		`access denied`		The `fruit` column was explicitly named in the query. The column-level security applies. Access is denied.
`SELECT color FROM my_dataset.my_table`			`access denied`	The `color` column was explicitly named in the query. The column-level security applies, before the row-level access policy on data in the `color` column is engaged. Access is denied.
`SELECT rank, fruit FROM my_dataset.my_table`		`access denied`		The ``fruit`` column was explicitly named in the query. The column-level security applies, before the row-level access policy on data in the `rank` column is engaged. Access is denied.
`SELECT rank, color FROM my_dataset.my_table`			`access denied`	The `color` column was explicitly named in the query The column-level security on the `color` column applies, before row-level access policies on data in the `rank` and the `color` columns are engaged. Access is denied.
`SELECT fruit, color FROM my_dataset.my_table`		`access denied`	`access denied`	The `fruit` and `color` columns were explicitly named in the query. The column-level security on the `fruit` and `color` columns applies, before the row-level access policy on data in the `color` column is engaged. Access is denied.
`SELECT * FROM my_dataset.my_table`		`access denied`	`access denied`	The `fruit` and `color` columns were implicitly named by using "`SELECT *`" in the query. The column-level security on the `fruit` and the `color` columns applies, before the row-level access policies on data in the `rank` or the `color` columns are engaged. Access is denied.

`TRUE` filter access

Finally, as explained in the section about TRUE filter access, if Alice or Bob has TRUE filter access, then they can see all of the rows in the table, and use it in non-query jobs. However, if the table has column-level security, then it still applies and can affect the results.

Extract jobs

If a table has row-level access policies, then only the data that you can view is exported to Cloud Storage when you run an extract job.

Legacy SQL

Row-level access policies are not compatible with Legacy SQL. Queries over tables with row-level access policies must use GoogleSQL. Legacy SQL queries are rejected.

Partitioned and clustered tables

Row-level security does not participate in query pruning, which is a feature of partitioned tables.

While row-level security is compatible with partitioned and clustered tables, the row-level access policies that filter row data aren't applied during partition pruning. You can still use partition pruning on a table that uses row-level security by specifying a WHERE clause that operates on the partition column. Similarly, row-level access policies themselves don't create any performance benefits for queries against clustered tables, but they don't interfere with other filtering that you apply.

Query pruning is performed during the execution of row-level access policies using the filters with the policies.

Rename a table

You do not need TRUE filter access to rename a table with one or more row access policies on it. You can rename a table with a DDL statement.

As an alternative, you can also copy a table and give the destination table a different name. If the source table has a row-level access policy on it, see table copy jobs on this page for more information.

Streaming updates

To perform streaming table UPDATE or DELETE operations with change data capture, you must have TRUE filter access.

Time travel

Only a table administrator can access historical data for a table that has, or has previously had, row-level access policies. Other users get an access denied error if they use a time travel decorator on a table that has had row-level access. For more information, see Time travel and row-level access.

Views and materialized views

The data displayed in a view or a materialized view is filtered according to the underlying source table's row-level access policies.

In addition, when a materialized view is derived from an underlying table that has row-level access policies, then the query performance is the same as when you query the source table directly. In other words, if the source table has row-level security, you don't see the typical performance benefits of querying a materialized view versus querying the source table.

You can't reference views or materialized views in row-level access policies.

Wildcard queries

Wildcard queries against tables with row-level access policies fail with an INVALID_INPUT error.

What's next

For information about best practices for row-level access policies, see Best practices for row-level security in BigQuery.