dagster-deltalake integration reference#

This reference page provides information for working with dagster-deltalake features that are not covered as part of the Using Dagster with Delta Lake tutorial.

Selecting specific columns in a downstream asset
Storing partitioned assets
Storing tables in multiple schemas
Using the Delta Lake I/O manager with other I/O managers
Storing and loading PyArrow Tables or Polars DataFrames in Delta Lake
Configuring storage backends

Selecting specific columns in a downstream asset#

Sometimes you may not want to fetch an entire table as the input to a downstream asset. With the Delta Lake I/O manager, you can select specific columns to load by supplying metadata on the downstream asset.


import pandas as pd

from dagster import AssetIn, asset

# this example uses the iris_dataset asset from Step 2 of the Using Dagster with Delta Lake tutorial


@asset(
    ins={
        "iris_sepal": AssetIn(
            key="iris_dataset",
            metadata={"columns": ["sepal_length_cm", "sepal_width_cm"]},
        )
    }
)
def sepal_data(iris_sepal: pd.DataFrame) -> pd.DataFrame:
    iris_sepal["sepal_area_cm2"] = (
        iris_sepal["sepal_length_cm"] * iris_sepal["sepal_width_cm"]
    )
    return iris_sepal

In this example, we only use the columns containing sepal data from the iris_dataset table created in Step 2 of the Using Dagster with Delta Lake tutorial. To select specific columns, we can add metadata to the input asset. We do this in the metadata parameter of the AssetIn that loads the iris_dataset asset in the ins parameter. We supply the key columns with a list of names of the columns we want to fetch.

When Dagster materializes sepal_data and loads the iris_dataset asset using the Delta Lake I/O manager, it will only fetch the sepal_length_cm and sepal_width_cm columns of the iris/iris_dataset table and pass them to sepal_data as a Pandas DataFrame.

Storing partitioned assets#

The Delta Lake I/O manager supports storing and loading partitioned data. To correctly store and load data from the Delta table, the Delta Lake I/O manager needs to know which column contains the data defining the partition bounds. The Delta Lake I/O manager uses this information to construct the correct queries to select or replace the data.

In the following sections, we describe how the I/O manager constructs these queries for different types of partitions.

For partitioning to work, the partition dimension needs to be one of the partition columns defined on the Delta table. Tables created via the I/O manager will be configured accordingly.

Storing static partitioned assets#

To store static partitioned assets in your Delta Lake, specify partition_expr metadata on the asset to tell the Delta Lake I/O manager which column contains the partition data:


import pandas as pd

from dagster import StaticPartitionsDefinition, asset


@asset(
    partitions_def=StaticPartitionsDefinition(
        ["Iris-setosa", "Iris-virginica", "Iris-versicolor"]
    ),
    metadata={"partition_expr": "species"},
)
def iris_dataset_partitioned(context) -> pd.DataFrame:
    species = context.partition_key

    full_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",
        ],
    )

    return full_df[full_df["species"] == species]


@asset
def iris_cleaned(iris_dataset_partitioned: pd.DataFrame):
    return iris_dataset_partitioned.dropna().drop_duplicates()

Dagster uses the partition_expr metadata to generate appropriate function parameters when loading the partition in the downstream asset. When loading a static partition this roughly corresponds to the following SQL statement:


SELECT *
 WHERE [partition_expr] in ([selected partitions])

A partition must be selected when materializing the above assets, as described in the Materializing partitioned assets documentation. In this example, the query used when materializing the Iris-setosa partition of the above assets would be:


SELECT *
 WHERE species = 'Iris-setosa'

Storing time-partitioned assets#

Like static partitioned assets, you can specify partition_expr metadata on the asset to tell the Delta Lake I/O manager which column contains the partition data:


import pandas as pd

from dagster import DailyPartitionsDefinition, asset


@asset(
    partitions_def=DailyPartitionsDefinition(start_date="2023-01-01"),
    metadata={"partition_expr": "time"},
)
def iris_data_per_day(context) -> pd.DataFrame:
    partition = context.partition_key

    # get_iris_data_for_date fetches all of the iris data for a given date,
    # the returned dataframe contains a column named 'time' with that stores
    # the time of the row as an integer of seconds since epoch
    return get_iris_data_for_date(partition)


@asset
def iris_cleaned(iris_data_per_day: pd.DataFrame):
    return iris_data_per_day.dropna().drop_duplicates()

Dagster uses the partition_expr metadata to craft the SELECT statement when loading the correct partition in the downstream asset. When loading a dynamic partition, the following statement is used:


SELECT *
 WHERE [partition_expr] = [partition_start]

A partition must be selected when materializing the above assets, as described in the Materializing partitioned assets documentation. The [partition_start] and [partition_end] bounds are of the form YYYY-MM-DD HH:MM:SS. In this example, the query when materializing the 2023-01-02 partition of the above assets would be:


SELECT *
 WHERE time = '2023-01-02 00:00:00'

Storing multi-partitioned assets#

The Delta Lake I/O manager can also store data partitioned on multiple dimensions. To do this, specify the column for each partition as a dictionary of partition_expr metadata:


import pandas as pd

from dagster import (
    DailyPartitionsDefinition,
    MultiPartitionsDefinition,
    StaticPartitionDefinition,
    asset,
)


@asset(
    partitions_def=MultiPartitionsDefinition(
        {
            "date": DailyPartitionsDefinition(start_date="2023-01-01"),
            "species": StaticPartitionDefinition(
                ["Iris-setosa", "Iris-virginica", "Iris-versicolor"]
            ),
        }
    ),
    metadata={"partition_expr": {"date": "time", "species": "species"}},
)
def iris_dataset_partitioned(context) -> pd.DataFrame:
    partition = context.partition_key.keys_by_dimension
    species = partition["species"]
    date = partition["date"]

    # get_iris_data_for_date fetches all of the iris data for a given date,
    # the returned dataframe contains a column named 'time' with that stores
    # the time of the row as an integer of seconds since epoch
    full_df = get_iris_data_for_date(date)

    return full_df[full_df["species"] == species]


@asset
def iris_cleaned(iris_dataset_partitioned: pd.DataFrame):
    return iris_dataset_partitioned.dropna().drop_duplicates()

Dagster uses the partition_expr metadata to craft the SELECT statement when loading the correct partition in a downstream asset. For multi-partitions, Dagster concatenates the WHERE statements described in the above sections to craft the correct SELECT statement.

A partition must be selected when materializing the above assets, as described in the Materializing partitioned assets documentation. For example, when materializing the 2023-01-02|Iris-setosa partition of the above assets, the following query will be used:


SELECT *
 WHERE species = 'Iris-setosa'
   AND time = '2023-01-02 00:00:00'

Storing tables in multiple schemas#

You may want to have different assets stored in different Delta Lake schemas. The Delta Lake I/O manager allows you to specify the schema in several ways.

If you want all of your assets to be stored in the same schema, you can specify the schema as configuration to the I/O manager, as we did in Step 1 of the Using Dagster with Delta Lake tutorial.

If you want to store assets in different schemas, you can specify the schema as part of the asset's asset key:

For SourceAsset, use the key parameter. The schema should be the second-to-last value in the parameter. In the following example, this would be daffodil.
For asset definitions, use the key_prefix parameter. This value will be prepended to the asset name to create the full asset key. In the following example, this would be iris.


import pandas as pd

from dagster import SourceAsset, asset

daffodil_dataset = SourceAsset(key=["daffodil", "daffodil_dataset"])


@asset(key_prefix=["iris"])
def iris_dataset() -> 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",
        ],
    )

In this example, the iris_dataset asset will be stored in the IRIS schema, and the daffodil_dataset asset will be found in the DAFFODIL schema.

The two options for specifying schema are mutually exclusive. If you provide schema configuration to the I/O manager, you cannot also provide it via the asset key and vice versa. If no schema is provided, either from configuration or asset keys, the default schema public will be used.

Using the Delta Lake I/O manager with other I/O managers#

You may have assets that you don't want to store in Delta Lake. You can provide an I/O manager to each asset using the io_manager_key parameter in the @asset decorator:


import pandas as pd
from dagster_aws.s3.io_manager import s3_pickle_io_manager
from dagster_deltalake import LocalConfig
from dagster_deltalake_pandas import DeltaLakePandasIOManager

from dagster import Definitions, asset


@asset(io_manager_key="warehouse_io_manager")
def iris_dataset() -> 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(io_manager_key="blob_io_manager")
def iris_plots(iris_dataset):
    # plot_data is a function we've defined somewhere else
    # that plots the data in a DataFrame
    return plot_data(iris_dataset)


defs = Definitions(
    assets=[iris_dataset, iris_plots],
    resources={
        "warehouse_io_manager": DeltaLakePandasIOManager(
            root_uri="path/to/deltalalke",
            storage_options=LocalConfig(),
            schema="iris",
        ),
        "blob_io_manager": s3_pickle_io_manager,
    },
)

In this example:

The iris_dataset asset uses the I/O manager bound to the key warehouse_io_manager and iris_plots uses the I/O manager bound to the key blob_io_manager
In the Definitions object, we supply the I/O managers for those keys
When the assets are materialized, the iris_dataset will be stored in Delta Lake, and iris_plots will be saved in Amazon S3

Storing and loading PyArrow tables or Polars DataFrames in Delta Lake#

The Delta Lake I/O manager also supports storing and loading PyArrow and Polars DataFrames.

Storing and loading PyArrow Tables with Delta Lake#

The deltalake package relies heavily on Apache Arrow for efficient data transfer, so PyArrow is natively supported.

You can use the DeltaLakePyArrowIOManager in a Definitions object as in Step 1 of the Using Dagster with Delta Lake tutorial.


from dagster_deltalake import DeltaLakePyarrowIOManager, LocalConfig

from dagster import Definitions

defs = Definitions(
    assets=[iris_dataset],
    resources={
        "io_manager": DeltaLakePyarrowIOManager(
            root_uri="path/to/deltalake",  # required
            storage_options=LocalConfig(),  # required
            schema="iris",  # optional, defaults to PUBLIC
        )
    },
)

Configuring storage backends#

The deltalake library comes with support for many storage backends out fo the box. WHich exact storage is to be used, is derived from the URL of a storage location.

S3 compatible storages#

The S3 APIs are implemented by a number of providers and it is possible to interact with many of them. However, most S3 implementations do not offer support for atomic operations, which is a requirement for multi writer support. As such some additional setup and configuration is required.

In case there will always be only a single writer to a table - this includes no concurrent dagster jobs writing to the same table - you can allow unsafe writes to the table.


from dagster_deltalake import S3Config

config = S3Config(allow_unsafe_rename=True)

To use DynamoDB, set the AWS_S3_LOCKING_PROVIDER variable to dynamodb and create a table named delta_rs_lock_table in Dynamo. An example DynamoDB table creation snippet using the aws CLI follows, and should be customized for your environment’s needs (e.g. read/write capacity modes):


aws dynamodb create-table --table-name delta_rs_lock_table \
    --attribute-definitions \
        AttributeName=key,AttributeType=S \
    --key-schema \
        AttributeName=key,KeyType=HASH \
    --provisioned-throughput \
        ReadCapacityUnits=10,WriteCapacityUnits=10

The delta-rs community is actively working on extending the awailable options for locking backends. This includes locking backends compatible with Databricks to allow concurrent writes from Databrticks and external environments.

Couldflare R2 storage has build in support for atomic copy operations. This can be leveraged by sending additional headers with the copy requrests.


from dagster_deltalake import S3Config

config = S3Config(copy_if_not_exists="header: cf-copy-destination-if-none-match: *")

In caeses where now-AWS S3 implementations are used, the endpoint URL or the S§ service needs to be provided.


config = S3Config(endpoint="https://<my-s3-endpoint-url>")

Working with locally running storage (emulators)#

A common pattern for e.g. integration tests is to run a storage emulator like Azurite, Localstack, o.a. If not configures to use TLS, we need to configure the http client, to allow for http traffic.


config = AzureConfig(use_emulator=True, client=ClientConfig(allow_http=True))