Using Dagster with Google BigQuery#

This tutorial focuses on creating and interacting with BigQuery tables using Dagster's asset definitions.

The dagster-gcp library provides two ways to interact with BigQuery tables:

• Resource: The resource allows you to directly run SQL queries against tables within an asset's compute function. Available resources: BigQueryResource
• I/O manager: The I/O manager transfers the responsibility of storing and loading DataFrames as BigQuery tables to Dagster. Available I/O managers: BigQueryPandasIOManager, BigQueryPySparkIOManager

This tutorial is divided into two sections to demonstrate the differences between the BigQuery resource and the BigQuery I/O manager. Each section will create the same assets, but the first section will use the BigQuery resource to store data in BigQuery, whereas the second section will use the BigQuery I/O manager. When writing your own assets, you may choose one or the other (or both) approaches depending on your storage requirements. See When to use I/O managers to learn more about when to use I/O managers and when to use resources.

In Option 1 you will:

• Set up and configure the BigQuery resource.
• Use the BigQuery resource to execute a SQL query to create a table.
• Use the BigQuery resource to execute a SQL query to interact with the table.

In Option 2 you will:

• Set up and configure the BigQuery I/O manager.
• Use Pandas to create a DataFrame, then delegate responsibility creating a table to the BigQuery I/O manager.
• Use the BigQuery I/O manager to load the table into memory so that you can interact with it using the Pandas library.

By the end of the tutorial, you will:

• Understand how to interact with a BigQuery database using the BigQuery resource.
• Understand how to use the BigQuery I/O manager to store and load DataFrames as BigQuery tables.
• Understand how to define dependencies between assets corresponding to tables in a BigQuery database.

Prerequisites#

To complete this tutorial, you'll need:

• To install the dagster-gcp and dagster-gcp-pandas libraries:

  pip install dagster-gcp dagster-gcp-pandas
  

• To gather the following information:

  • Google Cloud Project (GCP) project name: You can find this by logging into GCP and choosing one of the project names listed in the dropdown in the top left corner.

  • GCP credentials: You can authenticate with GCP two ways: by following GCP authentication instructions here, or by providing credentials directly to the BigQuery I/O manager.

    In this guide, we assume that you have run one of the gcloud auth commands or have set GOOGLE_APPLICATION_CREDENTIALS as specified in the linked instructions. For more information on providing credentials directly to the BigQuery resource and I/O manager, see Providing credentials as configuration in the BigQuery reference guide.

Option 1: Using the BigQuery resource#

Step 1: Configure the BigQuery resource#

To use the BigQuery resource, you'll need to add it to your Definitions object. The BigQuery resource requires some configuration:

• A project
• One method of authentication. You can follow the GCP authentication instructions here, or see Providing credentials as configuration in the BigQuery reference guide.

You can also specify a location where computation should take place.

from dagster_gcp import BigQueryResource

from dagster import Definitions

defs = Definitions(
    assets=[iris_data],
    resources={
        "bigquery": BigQueryResource(
            project="my-gcp-project",  # required
            location="us-east5",  # optional, defaults to the default location for the project - see https://cloud.google.com/bigquery/docs/locations for a list of locations
        )
    },
)

Step 2: Create tables in BigQuery#

import pandas as pd
from dagster_gcp import BigQueryResource

from dagster import asset


@asset
def iris_data(bigquery: BigQueryResource) -> None:
    iris_df = pd.read_csv(
        "https://docs.dagster.io/assets/iris.csv",
        names=[
            "sepal_length_cm",
            "sepal_width_cm",
            "petal_length_cm",
            "petal_width_cm",
            "species",
        ],
    )

    with bigquery.get_client() as client:
        job = client.load_table_from_dataframe(
            dataframe=iris_df,
            destination="iris.iris_data",
        )
        job.result()

from dagster import AssetSpec

iris_harvest_data = AssetSpec(key="iris_harvest_data")

Step 3: Define downstream assets#

Once you have created an asset that represents a table in BigQuery, you will likely want to create additional assets that work with the data.

from dagster import asset

from .create_table import iris_data

# this example uses the iris_dataset asset from Step 2


@asset(deps=[iris_data])
def iris_setosa(bigquery: BigQueryResource) -> None:
    job_config = bq.QueryJobConfig(destination="iris.iris_setosa")
    sql = "SELECT * FROM iris.iris_data WHERE species = 'Iris-setosa'"

    with bigquery.get_client() as client:
        job = client.query(sql, job_config=job_config)
        job.result()

In this asset, you're creating second table that only contains the data for the Iris Setosa species. This asset has a dependency on the iris_data asset. To define this dependency, you provide the iris_data asset as the deps parameter to the iris_setosa asset. You can then run the SQL query to create the table of Iris Setosa data.

Completed code example#

When finished, your code should look like the following:

import pandas as pd
from dagster_gcp import BigQueryResource
from google.cloud import bigquery as bq

from dagster import AssetSpec, Definitions, asset

iris_harvest_data = AssetSpec(key="iris_harvest_data")


@asset
def iris_data(bigquery: BigQueryResource) -> None:
    iris_df = pd.read_csv(
        "https://docs.dagster.io/assets/iris.csv",
        names=[
            "sepal_length_cm",
            "sepal_width_cm",
            "petal_length_cm",
            "petal_width_cm",
            "species",
        ],
    )

    with bigquery.get_client() as client:
        job = client.load_table_from_dataframe(
            dataframe=iris_df,
            destination="iris.iris_data",
        )
        job.result()


@asset(deps=[iris_data])
def iris_setosa(bigquery: BigQueryResource) -> None:
    job_config = bq.QueryJobConfig(destination="iris.iris_setosa")
    sql = "SELECT * FROM iris.iris_data WHERE species = 'Iris-setosa'"

    with bigquery.get_client() as client:
        job = client.query(sql, job_config=job_config)
        job.result()


defs = Definitions(
    assets=[iris_data, iris_setosa, iris_harvest_data],
    resources={
        "bigquery": BigQueryResource(
            project="my-gcp-project",
            location="us-east5",
        )
    },
)

Option 2: Using the BigQuery I/O manager#

You may want to use an I/O manager to handle storing DataFrames as tables in BigQuery and loading BigQuery tables as DataFrames in downstream assets. You may want to use an I/O manager if:

• You want your data to be loaded in memory so that you can interact with it using Python.
• You'd like to have Dagster manage how you store the data and load it as an input in downstream assets.

Using an I/O manager is not required, and you can reference When to use I/O managers to learn more.

This section of the guide focuses on storing and loading Pandas DataFrames in BigQuery, but Dagster also supports using PySpark DataFrames with BigQuery. The concepts from this guide apply to working with PySpark DataFrames, and you can learn more about setting up and using the BigQuery I/O manager with PySpark DataFrames in the reference guide.

Step 1: Configure the BigQuery I/O manager#

To use the BigQuery I/O manager, you'll need to add it to your Definitions object. The BigQuery I/O manager requires some configuration to connect to your Bigquery instance:

• A project
• One method of authentication. You can follow the GCP authentication instructions here, or see Providing credentials as configuration in the BigQuery reference guide.

You can also specify a location where data should be stored and processed and dataset that should hold the created tables. You can also set a timeout when working with Pandas DataFrames.

from dagster_gcp_pandas import BigQueryPandasIOManager

from dagster import Definitions

defs = Definitions(
    assets=[iris_data],
    resources={
        "io_manager": BigQueryPandasIOManager(
            project="my-gcp-project",  # required
            location="us-east5",  # optional, defaults to the default location for the project - see https://cloud.google.com/bigquery/docs/locations for a list of locations
            dataset="IRIS",  # optional, defaults to PUBLIC
            timeout=15.0,  # optional, defaults to None
        )
    },
)

With this configuration, if you materialized an asset called iris_data, the BigQuery I/O manager would store the data in the IRIS.IRIS_DATA table in the my-gcp-project project. The BigQuery instance would be located in us-east5.

Finally, in the Definitions object, we assign the BigQueryPandasIOManager to the io_manager key. io_manager is a reserved key to set the default I/O manager for your assets.

For more info about each of the configuration values, refer to the BigQueryPandasIOManager API documentation.

Step 2: Create tables in BigQuery#

The BigQuery I/O manager can create and update tables for your Dagster defined assets, but you can also make existing BigQuery tables available to Dagster.

import pandas as pd

from dagster import asset


@asset
def iris_data() -> pd.DataFrame:
    return pd.read_csv(
        "https://docs.dagster.io/assets/iris.csv",
        names=[
            "sepal_length_cm",
            "sepal_width_cm",
            "petal_length_cm",
            "petal_width_cm",
            "species",
        ],
    )

from dagster import AssetSpec

iris_harvest_data = AssetSpec(key="iris_harvest_data")

Step 3: Load BigQuery tables in downstream assets#

Once you have created an asset that represents a table in BigQuery, you will likely want to create additional assets that work with the data. Dagster and the BigQuery I/O manager allow you to load the data stored in BigQuery tables into downstream assets.

import pandas as pd

from dagster import asset

# this example uses the iris_data asset from Step 2


@asset
def iris_setosa(iris_data: pd.DataFrame) -> pd.DataFrame:
    return iris_data[iris_data["species"] == "Iris-setosa"]

In this asset, you're providing the iris_data asset from the Store a Dagster asset as a table in BigQuery example to the iris_setosa asset.

In this asset, you're providing the iris_data asset as a dependency to iris_setosa. By supplying iris_data as a parameter to iris_setosa, Dagster knows to use the BigQueryPandasIOManager to load this asset into memory as a Pandas DataFrame and pass it as an argument to iris_setosa. Next, a DataFrame that only contains the data for the Iris Setosa species is created and returned. Then the BigQueryPandasIOManager will store the DataFrame as the IRIS.IRIS_SETOSA table in BigQuery.

Completed code example#

When finished, your code should look like the following:

import pandas as pd
from dagster_gcp_pandas import BigQueryPandasIOManager

from dagster import AssetSpec, Definitions, asset

iris_harvest_data = AssetSpec(key="iris_harvest_data")


@asset
def iris_data() -> pd.DataFrame:
    return pd.read_csv(
        "https://docs.dagster.io/assets/iris.csv",
        names=[
            "sepal_length_cm",
            "sepal_width_cm",
            "petal_length_cm",
            "petal_width_cm",
            "species",
        ],
    )


@asset
def iris_setosa(iris_data: pd.DataFrame) -> pd.DataFrame:
    return iris_data[iris_data["species"] == "Iris-setosa"]


defs = Definitions(
    assets=[iris_data, iris_harvest_data, iris_setosa],
    resources={
        "io_manager": BigQueryPandasIOManager(
            project="my-gcp-project",
            location="us-east5",
            dataset="IRIS",
            timeout=15.0,
        )
    },
)

Related#

For more BigQuery features, refer to the BigQuery reference.

For more information on asset definitions, refer to the tutorial or the Assets concept documentation.

For more information on I/O managers, refer to the I/O manager concept documentation.