Transform data with data manipulation language (DML)

GoogleSQL is the new name for Google Standard SQL! New name, same great SQL dialect.

The BigQuery data manipulation language (DML) enables you to update, insert, and delete data from your BigQuery tables.

You can execute DML statements just as you would a SELECT statement, with the following conditions:

• You must use GoogleSQL. To enable GoogleSQL, see Switching SQL dialects.
• You cannot specify a destination table for the query.

For a list of BigQuery DML statements and examples of how to use them, see Data manipulation language statements in GoogleSQL. For more information about how to compute the number of bytes processed by a DML statement, see On-demand query size calculation.

Concurrent jobs

BigQuery manages the concurrency of DML statements that add, modify, or delete rows in a table.

Note: DML statements are subject to rate limits such as the maximum rate of table writes. You might hit a rate limit if you submit a high number of jobs against a table at one time. These rates do not limit the total number of DML statements that can be run. If you get an error message that says you've exceeded a rate limit, retry the operation using exponential backoff between retries.

INSERT DML concurrency

During any 24 hour period, the first 1500 INSERT statements run immediately after they are submitted. After this limit is reached, the concurrency of INSERT statements that write to a table is limited to 10. Additional INSERT statements are added to a PENDING queue. Up to 100 INSERT statements can be queued against a table at any given time. When an INSERT statement completes, the next INSERT statement is removed from the queue and run.

If you must run DML INSERT statements more frequently, consider streaming data to your table using the Storage Write API.

UPDATE, DELETE, MERGE DML concurrency

The UPDATE, DELETE, and MERGE DML statements are called mutating DML statements. If you submit one or more mutating DML statements on a table while other mutating DML jobs on it are still running (or pending), BigQuery runs up to 2 of them concurrently, after which up to 20 are queued as PENDING. When a previously running job finishes, the next pending job is dequeued and run. Queued mutating DML statements share a per-table queue with maximum length 20. Additional statements past the maximum queue length for each table fail with the error message: Resources exceeded during query execution: Too many DML statements outstanding against table PROJECT_ID:TABLE, limit is 20.

Interactive priority DML jobs that are queued for more than 6 hours fail with the following error message:

DML statement has been queued for too long

DML statement conflicts

Mutating DML statements that run concurrently on a table cause DML statement conflicts when the statements try to mutate the same partition. The statements succeed as long as they don't modify the same partition. BigQuery tries to rerun failed statements up to three times.

• An INSERT DML statement that inserts rows to a table doesn't conflict with any other concurrently running DML statement.

• A MERGE DML statement does not conflict with other concurrently running DML statements as long as the statement only inserts rows and does not delete or update any existing rows. This can include MERGE statements with UPDATE or DELETE clauses, as long as those clauses aren't invoked when the query runs.

Fine-grained DML

Preview

This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. You can process personal data for this feature as outlined in the Cloud Data Processing Addendum, subject to the obligations and restrictions described in the agreement under which you access Google Cloud. Pre-GA features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

Note: To provide feedback or request support for this feature, send an email to [e-mail].

Fine-grained DML is a performance enhancement designed to optimize the execution of UPDATE, DELETE, and MERGE statements (also known as mutating DML statements). Without fine-grained DML enabled, mutations are performed at the file-group level, which can lead to inefficient data rewrites. Fine-grained DML introduces a more granular approach that aims to reduce the amount of data that needs to be rewritten, and to reduce overall slot consumption.

To express interest in enrolling a project in the fine-grained DML preview, fill out the BigQuery fine-grained DML enrollment form. Projects are selectively enrolled based on an evaluation of your workloads.

Enable fine-grained DML

To enable fine-grained DML, set the enable_fine_grained_mutations table option to TRUE when you run a CREATE TABLE or ALTER TABLE DDL statement.

To create a new table with fine-grained DML, use the CREATE TABLE statement:

CREATE TABLE mydataset.mytable (
  product STRING,
  inventory INT64)
OPTIONS(enable_fine_grained_mutations = TRUE);

To alter an existing table with fine-grained DML, use the ALTER TABLE statement:

ALTER TABLE mydataset.mytable
SET OPTIONS(enable_fine_grained_mutations = TRUE);

After the enable_fine_grained_mutations option is set to TRUE, mutating DML statements are run with fine-grained DML capabilities enabled and use existing DML statement syntax.

To disable fine-grained DML on a table, set enable_fine_grained_mutations to FALSE by using the ALTER TABLE DDL statement.

Pricing

Enabling fine-grained DML for a table can incur additional BigQuery storage costs to store the extra mutation metadata that is associated with fine-grained DML operations. The actual cost depends on the amount of data that is modified, but for most situations it's expected to be negligible in comparison to the size of the table itself.

Projects configured to use reservations use slots to process fine-grained DML statements, including any background processing of table or mutation metadata.

Deleted data considerations

Fine-grained DML operations process deleted data in an offline manner.

Projects performing fine-grained DML operations without a BACKGROUND assignment process delete data using on-demand pricing. In this case, processing deleted data is performed regularly using internal BigQuery resources.

Projects performing fine-grained DML operations with a BACKGROUND assignment process deleted data using slots, and are subject to the configured reservation's resource availability. If there aren't enough resources available within the configured reservation, processing deleted data might take longer than anticipated.

Limitations

Tables enabled with fine-grained DML are subject to the following limitations:

• You can't use the tabledata.list method to read content from a table with fine-grained DML enabled. Instead, use the Storage Read API to read table records using an API.
• You can't create a table snapshot or table clone of a table with fine-grained DML enabled.
• 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.
• DML statements executed in a multi-statement transaction aren't optimized with fine-grained DML.

Best practices

For best performance, Google recommends the following patterns:

• Avoid submitting large numbers of individual row updates or insertions. Instead, group DML operations together when possible. For more information, see DML statements that update or insert single rows.

• If updates or deletions generally happen on older data, or within a particular range of dates, consider partitioning your tables. Partitioning ensures that the changes are limited to specific partitions within the table.

• Avoid partitioning tables if the amount of data in each partition is small and each update modifies a large fraction of the partitions.

• If you often update rows where one or more columns fall within a narrow range of values, consider using clustered tables. Clustering ensures that changes are limited to specific sets of blocks, reducing the amount of data that needs to be read and written. The following is an example of an UPDATE statement that filters on a range of column values:

  UPDATE mydataset.mytable
  SET string_col = 'some string'
  WHERE id BETWEEN 54 AND 75;
  

  Here is a similar example that filters on a small list of column values:

  UPDATE mydataset.mytable
  SET string_col = 'some string'
  WHERE id IN (54, 57, 60);
  

  Consider clustering on the id column in these cases.

• If you need OLTP functionality, consider using Cloud SQL federated queries, which enable BigQuery to query data that resides in Cloud SQL.

For best practices to optimize query performance, see Introduction to optimizing query performance.

Limitations

• Each DML statement initiates an implicit transaction, which means that changes made by the statement are automatically committed at the end of each successful DML statement.

• Rows that were recently written using the tabledata.insertall streaming method can't be modified with data manipulation language (DML), such as UPDATE, DELETE, MERGE, or TRUNCATE statements. The recent writes are those that occurred within the last 30 minutes. All other rows in the table remain modifiable by using UPDATE, DELETE, MERGE, or TRUNCATE statements. The streamed data can take up to 90 minutes to become available for copy operations.

  Alternatively, rows that were recently written using the Storage Write API can be modified using UPDATE, DELETE, or MERGE statements. For more information, see Use data manipulation language (DML) with recently streamed data.

• Correlated subqueries within a when_clause, search_condition, merge_update_clause or merge_insert_clause are not supported for MERGE statements.

• Queries that contain DML statements cannot use a wildcard table as the target of the query. For example, a wildcard table can be used in the FROM clause of an UPDATE query, but a wildcard table cannot be used as the target of the UPDATE operation.

What's next

• For DML syntax information and samples, see DML syntax.
• For information about using DML statements in scheduled queries, see Scheduling queries.