BigQuery configurations

Use `project` and `dataset` in configurations

schema is interchangeable with the BigQuery concept dataset
database is interchangeable with the BigQuery concept of project

For our reference documentation, you can declare project in place of database. This will allow you to read and write from multiple BigQuery projects. Same for dataset.

Using table partitioning and clustering

Partition clause

BigQuery supports the use of a partition by clause to easily partition a table by a column or expression. This option can help decrease latency and cost when querying large tables. Note that partition pruning only works when partitions are filtered using literal values (so selecting partitions using a subquery won't improve performance).

The partition_by config can be supplied as a dictionary with the following format:

{
  "field": "<field name>",
  "data_type": "<timestamp | date | datetime | int64>",
  "granularity": "<hour | day | month | year>"

  # Only required if data_type is "int64"
  "range": {
    "start": <int>,
    "end": <int>,
    "interval": <int>
  }
}

Partitioning by a date or timestamp

When using a datetime or timestamp column to partition data, you can create partitions with a granularity of hour, day, month, or year. A date column supports granularity of day, month and year. Daily partitioning is the default for all column types.

If the data_type is specified as a date and the granularity is day, dbt will supply the field as-is when configuring table partitioning.

Source code
Compiled code

bigquery_table.sql

{{ config(
    materialized='table',
    partition_by={
      "field": "created_at",
      "data_type": "timestamp",
      "granularity": "day"
    }
)}}

select
  user_id,
  event_name,
  created_at

from {{ ref('events') }}

bigquery_table.sql

create table `projectname`.`analytics`.`bigquery_table`
partition by timestamp_trunc(created_at, day)
as (

  select
    user_id,
    event_name,
    created_at

  from `analytics`.`events`

)

Partitioning by an "ingestion" date or timestamp

BigQuery supports an older mechanism of partitioning based on the time when each row was ingested. While we recommend using the newer and more ergonomic approach to partitioning whenever possible, for very large datasets, there can be some performance improvements to using this older, more mechanistic approach. Read more about the insert_overwrite incremental strategy below.

dbt will always instruct BigQuery to partition your table by the values of the column specified in partition_by.field. By configuring your model with partition_by.time_ingestion_partitioning set to True, dbt will use that column as the input to a _PARTITIONTIME pseudocolumn. Unlike with newer column-based partitioning, you must ensure that the values of your partitioning column match exactly the time-based granularity of your partitions.

Source code
Compiled code

bigquery_table.sql

{{ config(
    materialized="incremental",
    partition_by={
      "field": "created_date",
      "data_type": "timestamp",
      "granularity": "day",
      "time_ingestion_partitioning": true
    }
) }}

select
  user_id,
  event_name,
  created_at,
  -- values of this column must match the data type + granularity defined above
  timestamp_trunc(created_at, day) as created_date

from {{ ref('events') }}

bigquery_table.sql

create table `projectname`.`analytics`.`bigquery_table` (`user_id` INT64, `event_name` STRING, `created_at` TIMESTAMP)
partition by timestamp_trunc(_PARTITIONTIME, day);

insert into `projectname`.`analytics`.`bigquery_table` (_partitiontime, `user_id`, `event_name`, `created_at`)
select created_date as _partitiontime, * EXCEPT(created_date) from (
    select
      user_id,
      event_name,
      created_at,
      -- values of this column must match granularity defined above
      timestamp_trunc(created_at, day) as created_date

    from `projectname`.`analytics`.`events`
);

Partitioning with integer buckets

If the data_type is specified as int64, then a range key must also be provided in the partition_by dict. dbt will use the values provided in the range dict to generate the partitioning clause for the table.

Source code
Compiled code

bigquery_table.sql

{{ config(
    materialized='table',
    partition_by={
      "field": "user_id",
      "data_type": "int64",
      "range": {
        "start": 0,
        "end": 100,
        "interval": 10
      }
    }
)}}

select
  user_id,
  event_name,
  created_at

from {{ ref('events') }}

bigquery_table.sql

create table analytics.bigquery_table
partition by range_bucket(
  user_id,
  generate_array(0, 100, 10)
)
as (

  select
    user_id,
    event_name,
    created_at

  from analytics.events

)

Additional partition configs

If your model has partition_by configured, you may optionally specify two additional configurations:

require_partition_filter (boolean): If set to true, anyone querying this model must specify a partition filter, otherwise their query will fail. This is recommended for very large tables with obvious partitioning schemes, such as event streams grouped by day. Note that this will affect other dbt models or tests that try to select from this model, too.
partition_expiration_days (integer): If set for date- or timestamp-type partitions, the partition will expire that many days after the date it represents. E.g. A partition representing 2021-01-01, set to expire after 7 days, will no longer be queryable as of 2021-01-08, its storage costs zeroed out, and its contents will eventually be deleted. Note that table expiration will take precedence if specified.

bigquery_table.sql

{{ config(
    materialized = 'table',
    partition_by = {
      "field": "created_at",
      "data_type": "timestamp",
      "granularity": "day"
    },
    require_partition_filter = true,
    partition_expiration_days = 7
)}}

Clustering clause

BigQuery tables can be clustered to colocate related data.

Clustering on a single column:

bigquery_table.sql

{{
  config(
    materialized = "table",
    cluster_by = "order_id",
  )
}}

select * from ...

Clustering on multiple columns:

bigquery_table.sql

{{
  config(
    materialized = "table",
    cluster_by = ["customer_id", "order_id"],
  )
}}

select * from ...

Managing KMS encryption

Customer managed encryption keys can be configured for BigQuery tables using the kms_key_name model configuration.

Using KMS encryption

To specify the KMS key name for a model (or a group of models), use the kms_key_name model configuration. The following example sets the kms_key_name for all of the models in the encrypted/ directory of your dbt project.

dbt_project.yml

name: my_project
version: 1.0.0

...

models:
  my_project:
    encrypted:
      +kms_key_name: 'projects/PROJECT_ID/locations/global/keyRings/test/cryptoKeys/quickstart'

Labels and tags

Specifying labels

dbt supports the specification of BigQuery labels for the tables and views that it creates. These labels can be specified using the labels model config.

The labels config can be provided in a model config, or in the dbt_project.yml file, as shown below.

BigQuery key-value pair entries for labels larger than 63 characters are truncated.

Configuring labels in a model file

model.sql

{{
  config(
    materialized = "table",
    labels = {'contains_pii': 'yes', 'contains_pie': 'no'}
  )
}}

select * from {{ ref('another_model') }}

Configuring labels in dbt_project.yml

dbt_project.yml

models:
  my_project:
    snowplow:
      +labels:
        domain: clickstream
    finance:
      +labels:
        domain: finance

Viewing labels in the BigQuery console

Specifying tags

BigQuery table and view tags can be created by supplying an empty string for the label value.

model.sql

{{
  config(
    materialized = "table",
    labels = {'contains_pii': ''}
  )
}}

select * from {{ ref('another_model') }}

You can create a new label with no value or remove a value from an existing label key.

A label with a key that has an empty value can also be referred to as a tag in BigQuery. However, this should not be confused with a tag resource, which conditionally applies IAM policies to BigQuery tables and datasets. Find out more in labels and tags.

Currently, it's not possible to apply IAM tags in BigQuery, however, you can weigh in by upvoting GitHub issue 1134.

Policy tags

BigQuery enables column-level security by setting policy tags on specific columns.

dbt enables this feature as a column resource property, policy_tags (not a node config).

models/<filename>.yml

version: 2

models:
- name: policy_tag_table
  columns:
    - name: field
      policy_tags:
        - 'projects/<gcp-project>/locations/<location>/taxonomies/<taxonomy>/policyTags/<tag>'

Please note that in order for policy tags to take effect, column-level persist_docs must be enabled for the model, seed, or snapshot. Consider using variables to manage taxonomies and make sure to add the required security roles to your BigQuery service account key.

Merge behavior (incremental models)

The incremental_strategy config controls how dbt builds incremental models. dbt uses a merge statement on BigQuery to refresh incremental tables.

The incremental_strategy config can be set to one of the following values:

merge (default)
insert_overwrite
microbatch

Performance and cost

The operations performed by dbt while building a BigQuery incremental model can be made cheaper and faster by using a clustering clause in your model configuration. See this guide for more information on performance tuning for BigQuery incremental models.

Note: These performance and cost benefits are applicable to incremental models built with either the merge or the insert_overwrite incremental strategy.

The `merge` strategy

The merge incremental strategy will generate a merge statement that looks something like:

merge into {{ destination_table }} DEST
using ({{ model_sql }}) SRC
on SRC.{{ unique_key }} = DEST.{{ unique_key }}

when matched then update ...
when not matched then insert ...

The 'merge' approach automatically updates new data in the destination incremental table but requires scanning all source tables referenced in the model SQL, as well as destination tables. This can be slow and expensive for large data volumes. Partitioning and clustering techniques mentioned earlier can help mitigate these issues.

Note: The unique_key configuration is required when the merge incremental strategy is selected.

The `insert_overwrite` strategy

The insert_overwrite strategy generates a merge statement that replaces entire partitions in the destination table. Note: this configuration requires that the model is configured with a Partition clause. The merge statement that dbt generates when the insert_overwrite strategy is selected looks something like:

/*
  Create a temporary table from the model SQL
*/
create temporary table {{ model_name }}__dbt_tmp as (
  {{ model_sql }}
);

/*
  If applicable, determine the partitions to overwrite by
  querying the temp table.
*/

declare dbt_partitions_for_replacement array<date>;
set (dbt_partitions_for_replacement) = (
    select as struct
        array_agg(distinct date(max_tstamp))
    from `my_project`.`my_dataset`.{{ model_name }}__dbt_tmp
);

/*
  Overwrite partitions in the destination table which match
  the partitions in the temporary table
*/
merge into {{ destination_table }} DEST
using {{ model_name }}__dbt_tmp SRC
on FALSE

when not matched by source and {{ partition_column }} in unnest(dbt_partitions_for_replacement)
then delete

when not matched then insert ...

For a complete writeup on the mechanics of this approach, see this explainer post.

Determining partitions to overwrite

dbt is able to determine the partitions to overwrite dynamically from the values present in the temporary table, or statically using a user-supplied configuration.

The "dynamic" approach is simplest (and the default), but the "static" approach will reduce costs by eliminating multiple queries in the model build script.

Static partitions

To supply a static list of partitions to overwrite, use the partitions configuration.

models/session.sql

{% set partitions_to_replace = [
  'timestamp(current_date)',
  'timestamp(date_sub(current_date, interval 1 day))'
] %}

{{
  config(
    materialized = 'incremental',
    incremental_strategy = 'insert_overwrite',
    partition_by = {'field': 'session_start', 'data_type': 'timestamp'},
    partitions = partitions_to_replace
  )
}}

with events as (

    select * from {{ref('events')}}

    {% if is_incremental() %}
        -- recalculate yesterday + today
        where timestamp_trunc(event_timestamp, day) in ({{ partitions_to_replace | join(',') }})
    {% endif %}

),

... rest of model ...

This example model serves to replace the data in the destination table for both today and yesterday every day that it is run. It is the fastest and cheapest way to incrementally update a table using dbt. If we wanted this to run more dynamically— let’s say, always for the past 3 days—we could leverage dbt’s baked-in datetime macros and write a few of our own.

Think of this as "full control" mode. You must ensure that expressions or literal values in the partitions config have proper quoting when templated, and that they match the partition_by.data_type (timestamp, datetime, date, or int64). Otherwise, the filter in the incremental merge statement will raise an error.

Dynamic partitions

If no partitions configuration is provided, dbt will instead:

Create a temporary table for your model SQL
Query the temporary table to find the distinct partitions to be overwritten
Query the destination table to find the max partition in the database

When building your model SQL, you can take advantage of the introspection performed by dbt to filter for only new data. The maximum value in the partitioned field in the destination table will be available using the _dbt_max_partition BigQuery scripting variable. Note: this is a BigQuery SQL variable, not a dbt Jinja variable, so no jinja brackets are required to access this variable.

Example model SQL:

{{
  config(
    materialized = 'incremental',
    partition_by = {'field': 'session_start', 'data_type': 'timestamp'},
    incremental_strategy = 'insert_overwrite'
  )
}}

with events as (

  select * from {{ref('events')}}

  {% if is_incremental() %}

    -- recalculate latest day's data + previous
    -- NOTE: The _dbt_max_partition variable is used to introspect the destination table
    where date(event_timestamp) >= date_sub(date(_dbt_max_partition), interval 1 day)

{% endif %}

),

... rest of model ...

Copying partitions

If you are replacing entire partitions in your incremental runs, you can opt to do so with the copy table API and partition decorators rather than a merge statement. While this mechanism doesn't offer the same visibility and ease of debugging as the SQL merge statement, it can yield significant savings in time and cost for large datasets because the copy table API does not incur any costs for inserting the data - it's equivalent to the bq cp gcloud command line interface (CLI) command.

You can enable this by switching on copy_partitions: True in the partition_by configuration. This approach works only in combination with "dynamic" partition replacement.

bigquery_table.sql

{{ config(
    materialized="incremental",
    incremental_strategy="insert_overwrite",
    partition_by={
      "field": "created_date",
      "data_type": "timestamp",
      "granularity": "day",
      "time_ingestion_partitioning": true,
      "copy_partitions": true
    }
) }}

select
  user_id,
  event_name,
  created_at,
  -- values of this column must match the data type + granularity defined above
  timestamp_trunc(created_at, day) as created_date

from {{ ref('events') }}

logs/dbt.log

...
[0m16:03:13.017641 [debug] [Thread-3 (]: BigQuery adapter: Copying table(s) "/projects/projectname/datasets/analytics/tables/bigquery_table__dbt_tmp$20230112" to "/projects/projectname/datasets/analytics/tables/bigquery_table$20230112" with disposition: "WRITE_TRUNCATE"
...

Controlling table expiration

By default, dbt-created tables never expire. You can configure certain model(s) to expire after a set number of hours by setting hours_to_expiration.

Note

The hours_to_expiration only applies to initial creation of the underlying table. It doesn't reset for incremental models when they do another run.

dbt_project.yml

models:
  <resource-path>:
    +hours_to_expiration: 6

models/<modelname>.sql

{{ config(
    hours_to_expiration = 6
) }}

select ...

Authorized views

If the grant_access_to config is specified for a model materialized as a view, dbt will grant the view model access to select from the list of datasets provided. See BQ docs on authorized views for more details.

Note

The grants config and the grant_access_to config are distinct.

grant_access_to: Enables you to set up authorized views. When configured, dbt provides an authorized view access to show partial information from other datasets, without providing end users with full access to those underlying datasets. For more information, see "BigQuery configurations: Authorized views"
grants: Provides specific permissions to users, groups, or service accounts for managing access to datasets you're producing with dbt. For more information, see "Resource configs: grants"

You can use the two features together: "authorize" a view model with the grants_access_to configuration, and then add grants to that view model to share its query results (and only its query results) with other users, groups, or service accounts.

dbt_project.yml

models:
  <resource-path>:
    +grant_access_to:
      - project: project_1
        dataset: dataset_1
      - project: project_2
        dataset: dataset_2

models/<modelname>.sql

{{ config(
    grant_access_to=[
      {'project': 'project_1', 'dataset': 'dataset_1'},
      {'project': 'project_2', 'dataset': 'dataset_2'}
    ]
) }}

Views with this configuration will be able to select from objects in project_1.dataset_1 and project_2.dataset_2, even when they are located elsewhere and queried by users who do not otherwise have access to project_1.dataset_1 and project_2.dataset_2.

Materialized views

The BigQuery adapter supports materialized views with the following configuration parameters:

Parameter	Type	Required	Default	Change Monitoring Support
`on_configuration_change`	`<string>`	no	`apply`	n/a
`cluster_by`	`[<string>]`	no	`none`	drop/create
`partition_by`	`{<dictionary>}`	no	`none`	drop/create
`enable_refresh`	`<boolean>`	no	`true`	alter
`refresh_interval_minutes`	`<float>`	no	`30`	alter
`max_staleness` (in Preview)	`<interval>`	no	`none`	alter
`description`	`<string>`	no	`none`	alter
`labels`	`{<string>: <string>}`	no	`none`	alter
`hours_to_expiration`	`<integer>`	no	`none`	alter
`kms_key_name`	`<string>`	no	`none`	alter

Project file
Property file
Config block

dbt_project.yml

models:
  <resource-path>:
    +materialized: materialized_view
    +on_configuration_change: apply | continue | fail
    +cluster_by: <field-name> | [<field-name>]
    +partition_by:
      - field: <field-name>
      - data_type: timestamp | date | datetime | int64
        # only if `data_type` is not 'int64'
      - granularity: hour | day | month | year
        # only if `data_type` is 'int64'
      - range:
        - start: <integer>
        - end: <integer>
        - interval: <integer>
    +enable_refresh: true | false
    +refresh_interval_minutes: <float>
    +max_staleness: <interval>
    +description: <string>
    +labels: {<label-name>: <label-value>}
    +hours_to_expiration: <integer>
    +kms_key_name: <path-to-key>

models/properties.yml

version: 2

models:
  - name: [<model-name>]
    config:
      materialized: materialized_view
      on_configuration_change: apply | continue | fail
      cluster_by: <field-name> | [<field-name>]
      partition_by:
        - field: <field-name>
        - data_type: timestamp | date | datetime | int64
          # only if `data_type` is not 'int64'
        - granularity: hour | day | month | year
          # only if `data_type` is 'int64'
        - range:
          - start: <integer>
          - end: <integer>
          - interval: <integer>
      enable_refresh: true | false
      refresh_interval_minutes: <float>
      max_staleness: <interval>
      description: <string>
      labels: {<label-name>: <label-value>}
      hours_to_expiration: <integer>
      kms_key_name: <path-to-key>

models/<model_name>.sql

{{ config(
    materialized='materialized_view',
    on_configuration_change="apply" | "continue" | "fail",
    cluster_by="<field-name>" | ["<field-name>"],
    partition_by={
        "field": "<field-name>",
        "data_type": "timestamp" | "date" | "datetime" | "int64",

        # only if `data_type` is not 'int64'
        "granularity": "hour" | "day" | "month" | "year,

        # only if `data_type` is 'int64'
        "range": {
            "start": <integer>,
            "end": <integer>,
            "interval": <integer>,
        }
    },

    # auto-refresh options
    enable_refresh= true | false,
    refresh_interval_minutes=<float>,
    max_staleness="<interval>",

    # additional options
    description="<description>",
    labels={
        "<label-name>": "<label-value>",
    },
    hours_to_expiration=<integer>,
    kms_key_name="<path_to_key>",
) }}

Many of these parameters correspond to their table counterparts and have been linked above. The set of parameters unique to materialized views covers auto-refresh functionality.

Learn more about these parameters in BigQuery's docs:

Auto-refresh

Parameter	Type	Required	Default	Change Monitoring Support
`enable_refresh`	`<boolean>`	no	`true`	alter
`refresh_interval_minutes`	`<float>`	no	`30`	alter
`max_staleness` (in Preview)	`<interval>`	no	`none`	alter

BigQuery supports automatic refresh configuration for materialized views. By default, a materialized view will automatically refresh within 5 minutes of changes in the base table, but not more frequently than once every 30 minutes. BigQuery only officially supports the configuration of the frequency (the "once every 30 minutes" frequency); however, there is a feature in preview that allows for the configuration of the staleness (the "5 minutes" refresh). dbt will monitor these parameters for changes and apply them using an ALTER statement.

Learn more about these parameters in BigQuery's docs:

Limitations

As with most data platforms, there are limitations associated with materialized views. Some worth noting include:

Materialized view SQL has a limited feature set.
Materialized view SQL cannot be updated; the materialized view must go through a --full-refresh (DROP/CREATE).
The partition_by clause on a materialized view must match that of the underlying base table.
While materialized views can have descriptions, materialized view columns cannot.
Recreating/dropping the base table requires recreating/dropping the materialized view.

Find more information about materialized view limitations in Google's BigQuery docs.

Python model configuration

Submission methods: BigQuery supports a few different mechanisms to submit Python code, each with relative advantages. The dbt-bigquery adapter uses BigQuery DataFrames (BigFrames) or Dataproc. This process reads data from BigQuery, computes it either natively with BigQuery DataFrames or Dataproc, and writes the results back to BigQuery.

BigQuery DataFrames
Dataproc

BigQuery DataFrames can execute pandas and scikit-learn. There's no need to manage infrastructure and leverages BigQuery-distributed query engines. It's great for analysts, data scientists, and machine learning engineers who want to manipulate big data using a pandas-like syntax.

Note: BigQuery DataFrames run on Google Colab's default runtime. If no default runtime template is available, the adapter will automatically create one for you and mark it default for next time usage (assuming it has the right permissions).

BigQuery DataFrames setup:

# IAM permission if using service account

#Create Service Account
gcloud iam service-accounts create dbt-bigframes-sa
#Grant BigQuery User Role
gcloud projects add-iam-policy-binding ${GOOGLE_CLOUD_PROJECT} --member=serviceAccount:dbt-bigframes-sa@${GOOGLE_CLOUD_PROJECT}.iam.gserviceaccount.com --role=roles/bigquery.user
#Grant BigQuery Data Editor role. This can be restricted at dataset level
gcloud projects add-iam-policy-binding ${GOOGLE_CLOUD_PROJECT} --member=serviceAccount:dbt-bigframes-sa@${GOOGLE_CLOUD_PROJECT}.iam.gserviceaccount.com --role=roles/bigquery.dataEditor
#Grant Service Account user 
gcloud projects add-iam-policy-binding ${GOOGLE_CLOUD_PROJECT} --member=serviceAccount:dbt-bigframes-sa@${GOOGLE_CLOUD_PROJECT}.iam.gserviceaccount.com --role=roles/iam.serviceAccountUser
#Grant Colab Entperprise User
gcloud projects add-iam-policy-binding ${GOOGLE_CLOUD_PROJECT} --member=serviceAccount:dbt-bigframes-sa@${GOOGLE_CLOUD_PROJECT}.iam.gserviceaccount.com --role=roles/aiplatform.colabEnterpriseUser

# update the project.yaml file 
models:
  my_dbt_project:
    submission_method: bigframes


# update the profile.yaml file such as this
my_dbt_project_sa:
  outputs:
    dev:
      compute_region: us-central1
      dataset: <BIGQUERY_DATESET>
      gcs_bucket: <GCS BUCKET USED FOR BIGFRAME LOGS>
      job_execution_timeout_seconds: 300
      job_retries: 1
      keyfile: <SERVICE ACCOUNT KEY FILE>
      location: US
      method: service-account
      priority: interactive
      project: <BIGQUERY_PROJECT>
      threads: 1
      type: bigquery
  target: dev

Dataproc (serverless or pre-configured cluster) can execute Python models as PySpark jobs, reading from and writing to BigQuery. serverless is simpler but slower with limited configuration and pre-installed packages (pandas, numpy, scikit-learn), while cluster offers full control and faster runtimes. Good for complex, long-running batch pipelines and legacy Hadoop/Spark workflows but often slower for ad-hoc or interactive workloads.

Dataproc setup:

Create or use an existing Cloud Storage bucket.
Enable Dataproc APIs for your project and region.
If using the cluster submission method: Create or use an existing Dataproc cluster with the Spark BigQuery connector initialization action. (Google recommends copying the action into your own Cloud Storage bucket, rather than using the example version shown in the screenshot.)

Add the Spark BigQuery connector as an initialization action

The following configurations are needed to run Python models on Dataproc. You can add these to your BigQuery profile or configure them on specific Python models:

gcs_bucket: Storage bucket to which dbt will upload your model's compiled PySpark code.
dataproc_region: GCP region in which you have enabled Dataproc (for example us-central1).
dataproc_cluster_name: Name of Dataproc cluster to use for running Python model (executing PySpark job). Only required if submission_method: cluster.

def model(dbt, session):
    dbt.config(
        submission_method="cluster",
        dataproc_cluster_name="my-favorite-cluster"
    )
    ...

version: 2
models:
  - name: my_python_model
    config:
      submission_method: serverless

Python models running on Dataproc Serverless can be further configured in your BigQuery profile.

Any user or service account that runs dbt Python models will need the following permissions, in addition to the required BigQuery permissions:

dataproc.batches.create
dataproc.clusters.use
dataproc.jobs.create
dataproc.jobs.get
dataproc.operations.get
dataproc.operations.list
storage.buckets.get
storage.objects.create
storage.objects.delete

For more information, refer to Dataproc IAM roles and permissions.

Installing packages:

Installation of third-party packages on Dataproc varies depending on whether it's a cluster or serverless.

Dataproc Cluster — Google recommends installing Python packages while creating the cluster via initialization actions:
- How initialization actions are used
- Actions for installing via pip or conda
You can also install packages at cluster creation time by defining cluster properties: dataproc:pip.packages or dataproc:conda.packages.

Dataproc Serverless — Google recommends using a custom docker image to install thrid-party packages. The image needs to be hosted in Google Artifact Registry. It can then be used by providing the image path in dbt profiles:

profiles.yml

my-profile:
    target: dev
    outputs:
        dev:
        type: bigquery
        method: oauth
        project: abc-123
        dataset: my_dataset
        
        # for dbt Python models to be run on Dataproc Serverless
        gcs_bucket: dbt-python
        dataproc_region: us-central1
        submission_method: serverless
        dataproc_batch:
            runtime_config:
                container_image: {HOSTNAME}/{PROJECT_ID}/{IMAGE}:{TAG}

Adding packages to install via pip at cluster startup

Additional parameters

The BigQuery Python models also have the following additional configuration parameters:

Parameter	Type	Required	Default	Valid values
`enable_list_inference`	`<boolean>`	no	`True`	`True`, `False`
`intermediate_format`	`<string>`	no	`parquet`	`parquet`, `orc`
`submission_method`	`<string>`	no	``	`serverless`, `bigframes`, `cluster`
`notebook_template_id`	`<Integer>`	no	``	`<NOTEBOOK RUNTIME TEMPLATE_ID>`
`compute_region`	`<string>`	no	``	`<COMPUTE_REGION>`
`gcs_bucket`	`<string>`	no	``	`<GCS_BUCKET>`
`packages`	`<string>`	no	``	`['numpy<=1.1.1', 'pandas', 'mlflow']`

The enable_list_inference parameter
- The enable_list_inference parameter enables a PySpark data frame to read multiple records in the same operation. By default, this is set to True to support the default intermediate_format of parquet.
The intermediate_format parameter
- The intermediate_format parameter specifies which file format to use when writing records to a table. The default is parquet.
The submission_method parameter
- The submission_method parameter specifies whether the job will run on BigQuery DataFrames or Serverless Spark. submission_method is not required when dataproc_cluster_name is declared.
The notebook_template_id parameter
- The notebook_template_id parameter specifies runtime template in Colab Enterprise.
The compute_region parameter
- The compute_region parameter specifies the region of the job.
The gcs_bucket parameter
- The gcs_bucket parameter specifies the GCS bucket used for storing artifacts for the job.

Related docs:

Use project and dataset in configurations​

Using table partitioning and clustering​

Partition clause​

Partitioning by a date or timestamp​

Partitioning by an "ingestion" date or timestamp​

Partitioning with integer buckets​

Additional partition configs​

Clustering clause​

Managing KMS encryption​

Using KMS encryption​

Labels and tags​

Specifying labels​

Specifying tags​

Policy tags​

Merge behavior (incremental models)​

Performance and cost​

The merge strategy​

The insert_overwrite strategy​

Determining partitions to overwrite​

Static partitions​

Dynamic partitions​

Copying partitions​

Controlling table expiration​

Authorized views​

Materialized views​

Auto-refresh​

Limitations​

Python model configuration​

Additional parameters​

Use `project` and `dataset` in configurations

Using table partitioning and clustering

Partition clause

Partitioning by a date or timestamp

Partitioning by an "ingestion" date or timestamp

Partitioning with integer buckets

Additional partition configs

Clustering clause

Managing KMS encryption

Using KMS encryption

Labels and tags

Specifying labels

Specifying tags

Policy tags

Merge behavior (incremental models)

Performance and cost

The `merge` strategy

The `insert_overwrite` strategy

Determining partitions to overwrite

Static partitions

Dynamic partitions

Copying partitions

Controlling table expiration

Authorized views

Materialized views

Auto-refresh

Limitations

Python model configuration

Additional parameters