Cross-region dataset replication

With BigQuery dataset replication, you can set up automatic replication of a dataset between two different regions or multi-regions.

Overview

When you create a dataset in BigQuery, you select the region or multi-region where the data is stored. A region is a collection of data centers within a geographical area, and a multi-region is a large geographic area that contains two or more geographic regions. Your data is stored in one of the contained regions, and is not replicated within the multi-region. For more information about regions and multi-regions, see BigQuery locations.

BigQuery always stores copies of your data in two different Google Cloud zones within the dataset location. A zone is a deployment area for Google Cloud resources within a region. In all regions, replication between zones uses synchronous dual writes. Selecting a multi-region location does not provide cross-region replication or regional redundancy, so there is no increase in dataset availability in the event of a regional outage. Data is stored in a single region within the geographic location.

For additional geo-redundancy, you can replicate any dataset. BigQuery creates a secondary replica of the dataset, located in another region that you specify. This replica is then asynchronously replicated between two zones with the other region, for a total of four zonal copies.

Dataset replication

If you replicate a dataset, BigQuery stores the data in the region that you specify.

  • Primary region. When you first create a dataset, BigQuery places the dataset in the primary region.

  • Secondary region. When you add a dataset replica, BigQuery places the replica in the secondary region.

Initially, the replica in the primary region is the primary replica, and the replica in the secondary region is the secondary replica.

The primary replica is writeable, and the secondary replica is read-only. Writes to the primary replica are asynchronously replicated to the secondary replica. Within each region, the data is stored redundantly in two zones. Network traffic never leaves the Google Cloud network.

The following diagram shows the replication that occurs when a dataset is replicated:

The primary replica in the primary zone of region 1 is simultaneously replicated to the primary and secondary zones of region 2.

If the primary region is online, you can manually switch to the secondary replica. For more information, see Promote the secondary replica.

Pricing

You are billed for the following for replicated datasets:

  • Storage. Storage bytes in the secondary region are billed as a separate copy in the secondary region. See BigQuery storage pricing.
  • Data replication. For more information on how you are billed for data replication, see Data replication pricing.

Compute capacity in the secondary region

To run jobs and queries against the replica in the secondary region, you must purchase slots within the secondary region or run an on-demand query.

You can use the slots to perform read-only queries from the secondary replica. If you promote the secondary replica to be the primary, you can also use those slots to write to the replica.

You can purchase the same number of slots as you have in the primary region, or a different number of slots. If you purchase fewer slots, it might affect query performance.

Location considerations

Before adding a dataset replica, you need to create the initial dataset you want to replicate in BigQuery if it doesn't exist already. The location of the added replica is set to the location that you specify when adding the replica. The location of the added replica must be different from the location of the initial dataset. This means that the data in your dataset is continually replicated between the location the dataset was created in and the location of the replica. For replicas that require colocation, such as views, materialized views, or non-BigLake external tables, adding a replica in a location that is different from, or not compatible with, your source data's location could result in job errors.

When customers replicate a dataset across regions, BigQuery ensures data is located only in the locations where the replicas were created.

Colocation requirements

Using dataset replication is dependent on the following colocation requirements.

Cloud Storage

Querying data on Cloud Storage requires that the Cloud Storage bucket is colocated with the replica. Use the external tables location considerations when deciding where to place your replica.

Limitations

BigQuery dataset replication is subject to the following limitations:

  • Cross-region dataset replication is available within the same Google Cloud project. You can't create replicas if the source and target datasets are in different Google Cloud projects.
  • Data in streams that haven't been committed are not replicated in the secondary replica if the table is written to with the BigQuery Storage Write API. Replication of streaming data from the BigQuery Storage Write API or tabledata.insertAll is best-effort and may see high replication delay.
  • The BigQuery Storage Read API does not support reading from tables contained within the secondary dataset replica. To read from a table contained within the secondary replica, the secondary replica must first be promoted to be the primary replica.
  • Tables injected through Datastream to BigQuery by using Change Data Capture is not supported.
  • You can't enable fine-grained DML on a table in a replicated dataset, and you can't replicate a dataset that contains a table with fine-grained DML enabled.
  • Replication and switchover are managed through SQL data definition language (DDL) statements.
  • You are limited to one replica of each dataset for each region or multi-region. You cannot create two secondary replicas of the same dataset in the same destination region.
  • Resources within replicas are subject to the limitations as described in Resource behavior.
  • Policy tags and associated data policies are not replicated to the secondary replica. Any queries that reference columns with policy tags in regions other than the original region fail, even if that replica is promoted.
  • Time travel is only available in the secondary replica after the creation of the secondary replica is completed.
  • You can only replicate a dataset with fewer than 100,000 tables.
  • You are limited to a maximum of 4 replicas added (then dropped) to the same region per dataset per day.
  • You are limited by bandwidth.
  • Tables with Customer-managed encryption keys (CMEK) applied are not queryable in the secondary region if the replica_kms_key value is not configured.
  • BigLake tables are not supported.
  • You can't replicate external or federated datasets.
  • BigQuery Omni locations aren't supported.
  • You can't configure the following region pairs if you are configuring data replication for disaster recovery:
    • us-central1 - us multi-region
    • us-west1 - us multi-region
    • eu-west1 - eu multi-region
    • eu-west4 - eu multi-region

Resource behavior

The following operations are not supported on resources within the secondary replica:

If you need to create a copy of a resource in a secondary replica, you must copy the resource or query it and then materialize the results outside of the secondary replica. For example, use CREATE TABLE AS SELECT to create a new resource from the secondary replica resource.

Primary and secondary replicas are subject to the following differences:

Region 1 primary replica Region 2 secondary replica Notes
BigLake table BigLake table Not supported.
External table External table Only the external table definition is replicated. The query fails when the Cloud Storage bucket is not co-located in the same location as a replica.
Logical view Logical view Logical views that reference a dataset or resource that is not located in the same location as the logical view fail when queried.
Managed table Managed table No difference.
Materialized view Materialized view If a referenced table is not in the same region as the materialized view, the query fails. Replicated materialized views may see staleness above the view's max staleness.
Model Model Stored as managed tables.
Remote function Remote function Connections are regional. Remote functions that reference a dataset or resource (connection) that is not located in the same location as the remote function fail when run.
Routines User-defined function (UDF) or stored procedure Routines that reference a dataset or resource that is not located in the same location as the routine fail when run. Any routine that references a connection, such as remote functions, does not work outside the source region.
Row Access Policy Row Access Policy No difference.
Search index Search index Not replicated.
Stored procedure Stored procedure Stored procedures that reference a dataset or resource that is not located in the same location as the stored procedure fail when run.
Table clone Managed table Billed as a deep copy in secondary replica.
Table snapshot Table snapshot Billed as a deep copy in secondary replica.
Table-valued function (TVF) TVF TVFs that reference a dataset or resource that is not located in the same location as the TVF fail when run.
UDF UDF UDFs that reference a dataset or resource that is not located in the same location as the UDF fail when run.

Outage scenarios

Cross-region replication is not intended for use as a disaster recovery plan during a total-region outage. In the case of a total region outage in the primary replica's region, you cannot promote the secondary replica. Because secondary replicas are read-only, you can't run any write jobs on the secondary replica and can't promote the secondary region until the primary replica's region is restored. For more information about preparing for disaster recovery, see Managed disaster recovery.

The following table explains the impact of total-region outages on your replicated data:

Region 1 Region 2 Outage region Impact
Primary replica Secondary replica Region 2 Read-only jobs running in region 2 against the secondary replica fail.
Primary replica Secondary replica Region 1 All jobs running in region 1 fail. Read-only jobs continue to run in region 2 where the secondary replica is located. The contents of region 2 are stale until it is successfully synced with region 1.

Use dataset replication

This section describes how to replicate a dataset, promote the secondary replica, and run BigQuery read jobs in the secondary region.

Required permissions

To get the permissions that you need to manage replicas, ask your administrator to grant you bigquery.datasets.update permission.

Replicate a dataset

To replicate a dataset, use the ALTER SCHEMA ADD REPLICA DDL statement.

You can add a replica to any dataset that's located in a region or multi-region that is not already replicated in that region or multi-region. After you add a replica, it takes time for the initial copy operation to complete. You can still run queries referencing the primary replica while the data is being replicated, with no reduction in query processing capacity. You can't replicate data within the geo-locations within a multi-region.

The following example creates a dataset named my_dataset in the us-central1 region and then adds a replica in the us-east4 region:

-- Create the primary replica in the us-central1 region.
CREATE SCHEMA my_dataset OPTIONS(location='us-central1');

-- Create a replica in the secondary region.
ALTER SCHEMA my_dataset
ADD REPLICA `my_replica`
OPTIONS(location='us-east4');

To confirm when the secondary replica has successfully been created, you can query the creation_complete column in the INFORMATION_SCHEMA.SCHEMATA_REPLICAS view.

Promote the secondary replica

If the primary region is online, you can promote the secondary replica. Promotion switches the secondary replica to be the writeable primary. This operation completes within a few seconds if the secondary replica is caught up with the primary replica. If the secondary replica is not caught up, the promotion can't complete until it is caught up. The secondary replica can't be promoted to the primary if the region containing the primary has an outage.

Note the following:

  • All writes to tables return errors while promotion is in process. The old primary replica becomes non-writable immediately when the promotion begins.
  • Tables that aren't fully replicated at the time the promotion is initiated return stale reads.

To promote a replica to be the primary replica, use the ALTER SCHEMA SET OPTIONS DDL statement and set the primary_replica option.

Note the following: - You must explicitly set the job location to the secondary region in query settings. See BigQuery specify locations.

The following example promotes the us-east4 replica to be the primary:

ALTER SCHEMA my_dataset SET OPTIONS(primary_replica = 'us-east4')

To confirm when the secondary replica has successfully been promoted, you can query the replica_primary_assignment_complete column in the INFORMATION_SCHEMA.SCHEMATA_REPLICAS view.

Remove a dataset replica

To remove a replica and stop replicating the dataset, use the ALTER SCHEMA DROP REPLICA DDL statement.

The following example removes the us replica:

ALTER SCHEMA my_dataset
DROP REPLICA IF EXISTS `us`;

You must first drop any secondary replicas to delete the entire dataset. If you delete the entire dataset—for example, by using the DROP SCHEMAstatement—without dropping all secondary replicas, you receive an error.

List dataset replicas

To list the dataset replicas in a project, query the INFORMATION_SCHEMA.SCHEMATA_REPLICAS view.

Migrate datasets

You can use cross-region dataset replication to migrate your datasets from one region to another. The following example demonstrates the process of migrating the existing my_migration dataset from the US multi-region to the EU multi-region using cross-region replication.

Replicate the dataset

To begin the migration process, first replicate the dataset in the region that you want to migrate the data to. In this scenario, you are migrating the my_migration dataset to the EU multi-region.

-- Create a replica in the secondary region.
ALTER SCHEMA my_migration
ADD REPLICA `eu`
OPTIONS(location='eu');

This creates a secondary replica named eu in the EU multi-region. The primary replica is the my_migration dataset in the US multi-region.

Promote the secondary replica

To continue migrating the dataset to the EU multi-region, promote the secondary replica:

ALTER SCHEMA my_migration SET OPTIONS(primary_replica = 'eu')

After the promotion is complete, eu is the primary replica. It is a writable replica.

Complete the migration

To complete the migration from the US multi-region to the EU multi-region, delete the us replica. This step is not required but is useful if you don't need a dataset replica beyond your migration needs.

ALTER SCHEMA my_migration
DROP REPLICA IF EXISTS us;

Your dataset is located in the EU multi-region and there are no replicas of the my_migration dataset. You have successfully migrated your dataset to the EU multi-region. The complete list of resources that are migrated can be found in Resource behavior.

Customer-managed encryption keys (CMEK)

Customer-managed Cloud Key Management Service keys are not automatically replicated when you create a secondary replica. In order to maintain the encryption on your replicated dataset, you must set the replica_kms_key for the location of the added replica. You can set the replica_kms_key using the ALTER SCHEMA ADD REPLICA DDL statement.

Replicating datasets with CMEK behaves as described in the following scenarios:

  • If the source dataset has a default_kms_key, you must provide a replica_kms_key that was created in the replica dataset's region when using the ALTER SCHEMA ADD REPLICA DDL statement.

  • If the source dataset doesn't have a value set for default_kms_key, you can't set the replica_kms_key.

  • If you are using Cloud KMS key rotation on either (or both) of the default_kms_key or the replica_kms_key the replicated dataset is still queryable after the key rotation.

    • Key rotation in the primary region updates the key version only in tables created after the rotation, tables that existed prior to the key rotation still use the key version that was set prior to the rotation.
    • Key rotation in the secondary region updates all tables in the secondary replica to the new key version.
    • Switching the primary replica to secondary replica updates all tables in the secondary replica (formerly the primary replica) to the new key version.
    • If the key version set on tables in the primary replica prior to key rotation is deleted, any tables still using the key version set prior to key rotation cannot be queried until the key version is updated. In order to update the key version, the old key version must be active (not disabled or deleted).
  • If the source dataset doesn't have a value set for default_kms_key, but there are individual tables in the source dataset with CMEK applied, those tables aren't queryable in the replicated dataset. To query the tables, do the following:

    • Add a default_kms_key value for the source dataset.
    • When you create a new replica using the ALTER SCHEMA ADD REPLICA DDL statement, set a value for the replica_kms_keyoption. The CMEK tables are queryable in the destination region.

    All the CMEK tables in the destination region use the same replica_kms_key, regardless of the key used in the source region.

Create a replica with CMEK

The following example creates a replica in the us-west1 region with a replica_kms_key value set. For CMEK key, grant the BigQuery service account permission to encrypt and decrypt.

-- Create a replica in the secondary region.
ALTER SCHEMA my_dataset
ADD REPLICA `us-west1`
OPTIONS(location='us-west1',
  replica_kms_key='my_us_west1_kms_key_name');

CMEK limitations

Replicating datasets with CMEK applied are subject to the following limitations:

  • You can't update the replicated Cloud KMS key after the replica is created.

  • You can't update the default_kms_key value on the source dataset after the dataset replicas have been created.

  • If the provided replica_kms_key is not valid in the destination region, the dataset won't be replicated.

  • CMEK-encrypted tables replicated to the destination region are visible, but cannot be queried or written to because the source CMEK is not recognized in the destination region.

What's next