Google BigQuery#

Install dlt with BigQuery#

To install the dlt library with BigQuery dependencies:

pip install "dlt[bigquery]"

Setup guide#

1. Initialize a project with a pipeline that loads to BigQuery by running:

dlt init chess bigquery

2. Install the necessary dependencies for BigQuery by running:

pip install -r requirements.txt

This will install dlt with the bigquery extra, which contains all the dependencies required by the BigQuery client.

3. Log in to or create a Google Cloud account

Sign up for or log in to the Google Cloud Platform in your web browser.

4. Create a new Google Cloud project

After arriving at the Google Cloud console welcome page, click the project selector in the top left, then click the New Project button, and finally click the Create button after naming the project whatever you would like.

5. Create a service account and grant BigQuery permissions

You will then need to create a service account. After clicking the Go to Create service account button on the linked docs page, select the project you created and name the service account whatever you would like.

Click the Continue button and grant the following roles, so that dlt can create schemas and load data:

• BigQuery Data Editor
• BigQuery Job User
• BigQuery Read Session User

You don't need to grant users access to this service account now, so click the Done button.

6. Download the service account JSON

In the service accounts table page that you're redirected to after clicking Done as instructed above, select the three dots under the Actions column for the service account you created and select Manage keys.

This will take you to a page where you can click the Add key button, then the Create new key button, and finally the Create button, keeping the preselected JSON option.

A JSON file that includes your service account private key will then be downloaded.

7. Update your dlt credentials file with your service account info

Open your dlt credentials file:

open .dlt/secrets.toml

Replace the project_id, private_key, and client_email with the values from the downloaded JSON file:

[destination.bigquery]
location = "US"

[destination.bigquery.credentials]
project_id = "project_id" # please set me up!
private_key = "private_key" # please set me up!
client_email = "client_email" # please set me up!

You can specify the location of the data, i.e., EU instead of US, which is the default.

OAuth 2.0 authentication#

You can use OAuth 2.0 authentication. You'll need to generate a refresh token with the right scopes (we suggest asking our GPT-4 assistant for details). Then you can fill the following information in secrets.toml:

[destination.bigquery]
location = "US"

[destination.bigquery.credentials]
project_id="project_id" # please set me up!
client_id = "client_id" # please set me up!
client_secret = "client_secret" # please set me up!
refresh_token = "refresh_token" # please set me up!

Using default credentials#

Google provides several ways to get default credentials, i.e., from the GOOGLE_APPLICATION_CREDENTIALS environment variable or metadata services. VMs available on GCP (cloud functions, Composer runners, Colab notebooks) have associated service accounts or authenticated users. dlt will try to use default credentials if nothing is explicitly specified in the secrets.

[destination.bigquery]
location = "US"

Using different project_id#

You can set the project_id in your configuration to be different from the one in your credentials, provided your account has access to it:

[destination.bigquery]
project_id = "project_id_destination"

[destination.bigquery.credentials]
project_id = "project_id_credentials"

In this scenario, project_id_credentials will be used for authentication, while project_id_destination will be used as the data destination.

Write disposition#

All write dispositions are supported.

If you set the replace strategy to staging-optimized, the destination tables will be dropped and recreated with a clone command from the staging tables.

Data loading#

dlt uses BigQuery load jobs that send files from the local filesystem or GCS buckets.
The loader follows Google recommendations when retrying and terminating jobs.
The Google BigQuery client implements an elaborate retry mechanism and timeouts for queries and file uploads, which may be configured in destination options.

BigQuery destination also supports streaming insert. The mode provides better performance with small batches, but it buffers the data, preventing any update/delete operations on it. Due to this, streaming inserts are only available with write_disposition="append", and the inserted data is blocked for editing for up to 90 min (reading, however, is available immediately). See more.

To switch the resource into streaming insert mode, use hints:

@dlt.resource(write_disposition="append")
def streamed_resource():
    yield {"field1": 1, "field2": 2}

streamed_resource.apply_hints(additional_table_hints={"x-insert-api": "streaming"})

Use BigQuery schema autodetect for nested fields#

You can let BigQuery infer schemas and create destination tables instead of dlt. As a consequence, nested fields (i.e., RECORD), which dlt does not support at
this moment (they are stored as JSON), may be created. You can select certain resources with the BigQuery Adapter or all of them with the following config option:

[destination.bigquery]
autodetect_schema=true

We recommend yielding Arrow tables from your resources and using the Parquet file format to load the data. In that case, the schemas generated by dlt and BigQuery
will be identical. BigQuery will also preserve the column order from the generated parquet files. You can convert JSON data into Arrow tables with pyarrow or duckdb.

import pyarrow.json as paj

import dlt
from dlt.destinations.adapters import bigquery_adapter

@dlt.resource(name="cve")
def load_cve():
  with open("cve.json", 'rb') as f:
    # autodetect arrow schema and yield arrow table
    yield paj.read_json(f)

pipeline = dlt.pipeline("load_json_struct", destination="bigquery")
pipeline.run(
  bigquery_adapter(load_cve(), autodetect_schema=True)
)

Above, we use the pyarrow library to convert a JSON document into an Arrow table and use bigquery_adapter to enable schema autodetect for the cve resource.

Yielding Python dicts/lists and loading them as JSONL works as well. In many cases, the resulting nested structure is simpler than those obtained via pyarrow/duckdb and parquet. However, there are slight differences in inferred types from dlt (BigQuery coerces types more aggressively). BigQuery also does not try to preserve the column order in relation to the order of fields in JSON.

import dlt
from dlt.destinations.adapters import bigquery_adapter

@dlt.resource(name="cve", max_table_nesting=1)
def load_cve():
  with open("cve.json", 'rb') as f:
    yield json.load(f)

pipeline = dlt.pipeline("load_json_struct", destination="bigquery")
pipeline.run(
  bigquery_adapter(load_cve(), autodetect_schema=True)
)

In the example below, we represent JSON data as tables up to nesting level 1. Above this nesting level, we let BigQuery create nested fields.

Supported file formats#

You can configure the following file formats to load data to BigQuery:

• JSONL is used by default.
• Parquet is supported.

When staging is enabled:

• JSONL is used by default.
• Parquet is supported.

Supported column hints#

BigQuery supports the following column hints:

• partition - creates a partition with a day granularity on the decorated column (PARTITION BY DATE).
  It may be used with datetime, date, and bigint data types.
  Only one column per table is supported and only when a new table is created.
  For more information on BigQuery partitioning, read the official docs.

  ❗ bigint maps to BigQuery's INT64 data type.
  Automatic partitioning requires converting an INT64 column to a UNIX timestamp, which GENERATE_ARRAY doesn't natively support.
  With a 10,000 partition limit, we can’t cover the full INT64 range.
  Instead, we set 86,400-second boundaries to enable daily partitioning.
  This captures typical values, but extremely large/small outliers go to an __UNPARTITIONED__ catch-all partition.

• cluster - creates cluster column(s). Many columns per table are supported and only when a new table is created.

Table and column identifiers#

BigQuery uses case-sensitive identifiers by default, and this is what dlt assumes. If the dataset you use has case-insensitive identifiers (you have such an option
when you create it), make sure that you use a case-insensitive naming convention or you tell dlt about it so identifier collisions are properly detected.

[destination.bigquery]
has_case_sensitive_identifiers=false

You have an option to allow dlt to set the case sensitivity for newly created datasets. In that case, it will follow the case sensitivity of the current
naming convention (i.e., the default snake_case will create a dataset with case-insensitive identifiers).

[destination.bigquery]
should_set_case_sensitivity_on_new_dataset=true

The option above is off by default.

Staging support#

BigQuery supports GCS as a file staging destination. dlt will upload files in the parquet format to GCS and ask BigQuery to copy their data directly into the database.
Please refer to the Google Storage filesystem documentation to learn how to set up your GCS bucket with the bucket_url and credentials.
If you use the same service account for GCS and your BigQuery deployment, you do not need to provide additional authentication for BigQuery to be able to read from your bucket.

Alternatively to parquet files, you can specify jsonl as the staging file format. For this, set the loader_file_format argument of the run command of the pipeline to jsonl.

BigQuery/GCS staging example#

# Create a dlt pipeline that will load
# chess player data to the BigQuery destination
# via a GCS bucket.
pipeline = dlt.pipeline(
    pipeline_name='chess_pipeline',
    destination='bigquery',
    staging='filesystem', # Add this to activate the staging location.
    dataset_name='player_data'
)

Additional destination options#

You can configure the data location and various timeouts as shown below. This information is not a secret so it can be placed in config.toml as well:

[destination.bigquery]
location="US"
http_timeout=15.0
file_upload_timeout=1800.0
retry_deadline=60.0

• location sets the BigQuery data location (default: US)
• http_timeout sets the timeout when connecting and getting a response from the BigQuery API (default: 15 seconds)
• file_upload_timeout is a timeout for file upload when loading local files: the total time of the upload may not exceed this value (default: 30 minutes, set in seconds)
• retry_deadline is a deadline for a DEFAULT_RETRY used by Google
• ignore_unknown_values is a configuration option that allows BigQuery to ignore rows with unknown or unexpected values during data loading. When enabled, rows containing fields that are not defined in the schema will be skipped instead of causing the entire load job to fail. This can be useful when dealing with inconsistent or evolving data sources.

dbt support#

This destination integrates with dbt via dbt-bigquery.
Credentials, if explicitly defined, are shared with dbt along with other settings like location, retries, and timeouts.
In the case of implicit credentials (i.e., available in a cloud function), dlt shares the project_id and delegates obtaining credentials to the dbt adapter.

Syncing of dlt state#

This destination fully supports dlt state sync.

BigQuery adapter#

You can use the bigquery_adapter to add BigQuery-specific hints to a resource.
These hints influence how data is loaded into BigQuery tables, such as specifying partitioning, clustering, and numeric column rounding modes.
Hints can be defined at both the column level and table level.

The adapter updates the DltResource with metadata about the destination column and table DDL options.

Use an adapter to apply hints to a resource#

Here is an example of how to use the bigquery_adapter method to apply hints to a resource on both the column level and table level:

import dlt
from dlt.destinations.adapters import bigquery_adapter

@dlt.resource(
    columns=[
        {"name": "event_date", "data_type": "date"},
        {"name": "user_id", "data_type": "bigint"},
        # Other columns.
    ]
)
def event_data():
    yield from [
        {"event_date": datetime.date.today() + datetime.timedelta(days=i)} for i in range(100)
    ]

# Apply column options.
bigquery_adapter(
    event_data, partition="event_date", cluster=["event_date", "user_id"]
)

# Apply table level options.
bigquery_adapter(event_data, table_description="Dummy event data.")

# Load data in "streaming insert" mode (only available with
# write_disposition="append").
bigquery_adapter(event_data, insert_api="streaming")

In the example above, the adapter specifies that event_date should be used for partitioning and both event_date and user_id should be used for clustering (in the given order) when the table is created.

Some things to note with the adapter's behavior:

• You can only partition on one column (refer to supported hints).
• You can cluster on as many columns as you would like.
• Sequential adapter calls on the same resource accumulate parameters, akin to an OR operation, for a unified execution.