Execution Modes

Cosmos can run dbt commands using five different approaches, called execution modes:

1. local: Run dbt commands using a local dbt installation (default)

2. virtualenv: Run dbt commands from Python virtual environments managed by Cosmos

3. docker: Run dbt commands from Docker containers managed by Cosmos (requires a pre-existing Docker image)

4. kubernetes: Run dbt commands from Kubernetes Pods managed by Cosmos (requires a pre-existing Docker image)

5. aws_eks: Run dbt commands from AWS EKS Pods managed by Cosmos (requires a pre-existing Docker image)

6. azure_container_instance: Run dbt commands from Azure Container Instances managed by Cosmos (requires a pre-existing Docker image)

7. gcp_cloud_run_job: Run dbt commands from GCP Cloud Run Job instances managed by Cosmos (requires a pre-existing Docker image)

8. airflow_async: (Experimental and introduced since Cosmos 1.7.0) Run the dbt resources from your dbt project asynchronously, by submitting the corresponding compiled SQLs to Apache Airflow’s Deferrable operators

The choice of the execution mode can vary based on each user’s needs and concerns. For more details, check each execution mode described below.

Execution Modes Comparison

Execution Mode	Task Duration	Environment Isolation	Cosmos Profile Management
Local	Fast	None	Yes
Virtualenv	Medium	Lightweight	Yes
Docker	Slow	Medium	No
Kubernetes	Slow	High	No
AWS_EKS	Slow	High	No
Azure Container Instance	Slow	High	No
GCP Cloud Run Job Instance	Slow	High	No
Airflow Async	Medium	None	Yes

Local

By default, Cosmos uses the local execution mode.

The local execution mode is the fastest way to run Cosmos operators since they don’t install dbt nor build docker containers. However, it may not be an option for users using managed Airflow services such as Google Cloud Composer, since Airflow and dbt dependencies can conflict (Airflow and dbt dependencies conflicts), the user may not be able to install dbt in a custom path.

The local execution mode assumes a dbt binary is reachable within the Airflow worker node.

If dbt was not installed as part of the Cosmos packages, users can define a custom path to dbt by declaring the argument dbt_executable_path.

Note

Starting in the 1.4 version, Cosmos tries to leverage the dbt partial parsing (partial_parse.msgpack) to speed up task execution. This feature is bound to dbt partial parsing limitations. Learn more: Partial parsing.

When using the local execution mode, Cosmos converts Airflow Connections into a native dbt profiles file (profiles.yml).

Example of how to use, for instance, when dbt was installed together with Cosmos:

basic_cosmos_dag = DbtDag(
    # dbt/cosmos-specific parameters
    project_config=ProjectConfig(
        DBT_ROOT_PATH / "jaffle_shop",
    ),
    profile_config=profile_config,
    operator_args={
        "install_deps": True,  # install any necessary dependencies before running any dbt command
        "full_refresh": True,  # used only in dbt commands that support this flag
    },
    # normal dag parameters
    schedule_interval="@daily",
    start_date=datetime(2023, 1, 1),
    catchup=False,
    dag_id="basic_cosmos_dag",
    default_args={"retries": 2},
)

Virtualenv

If you’re using managed Airflow on GCP (Cloud Composer), for instance, we recommend you use the virtualenv execution mode.

The virtualenv mode isolates the Airflow worker dependencies from dbt by managing a Python virtual environment created during task execution and deleted afterwards.

In this case, users are responsible for declaring which version of dbt they want to use by giving the argument py_requirements. This argument can be set directly in operator instances or when instantiating DbtDag and DbtTaskGroup as part of operator_args.

Similar to the local execution mode, Cosmos converts Airflow Connections into a way dbt understands them by creating a dbt profile file (profiles.yml). Also similar to the local execution mode, Cosmos will by default attempt to use a partial_parse.msgpack if one exists to speed up parsing.

Some drawbacks of this approach:

• It is slower than local because it creates a new Python virtual environment for each Cosmos dbt task run.

• If dbt is unavailable in the Airflow scheduler, the default LoadMode.DBT_LS will not work. In this scenario, users must use a parsing method that does not rely on dbt, such as LoadMode.MANIFEST.

• Only InvocationMode.SUBPROCESS is supported currently, attempt to use InvocationMode.DBT_RUNNER will raise error.

Example of how to use:

@dag(
    schedule_interval="@daily",
    start_date=datetime(2023, 1, 1),
    catchup=False,
)
def example_virtualenv() -> None:
    start_task = EmptyOperator(task_id="start-venv-examples")
    end_task = EmptyOperator(task_id="end-venv-examples")

    # This first task group creates a new Cosmos virtualenv every time a task is run
    # and deletes it afterwards
    # It is much slower than if the user sets the `virtualenv_dir`
    tmp_venv_task_group = DbtTaskGroup(
        group_id="tmp-venv-group",
        # dbt/cosmos-specific parameters
        project_config=ProjectConfig(
            DBT_ROOT_PATH / "jaffle_shop",
        ),
        profile_config=profile_config,
        execution_config=ExecutionConfig(
            execution_mode=ExecutionMode.VIRTUALENV,
            # Without setting virtualenv_dir="/some/path/persistent-venv",
            # Cosmos creates a new Python virtualenv for each dbt task being executed
        ),
        operator_args={
            "py_system_site_packages": False,
            "py_requirements": ["dbt-postgres"],
            "install_deps": True,
            "emit_datasets": False,  # Example of how to not set inlets and outlets
        },
    )

    # The following task group reuses the Cosmos-managed Python virtualenv across multiple tasks.
    # It runs approximately 70% faster than the previous TaskGroup.
    cached_venv_task_group = DbtTaskGroup(
        group_id="cached-venv-group",
        # dbt/cosmos-specific parameters
        project_config=ProjectConfig(
            DBT_ROOT_PATH / "jaffle_shop",
        ),
        profile_config=profile_config,
        execution_config=ExecutionConfig(
            execution_mode=ExecutionMode.VIRTUALENV,
            # We can set the argument `virtualenv_dir` if we want Cosmos to create one Python virtualenv
            # and reuse that to run all the dbt tasks within the same worker node
            virtualenv_dir=Path("/tmp/persistent-venv2"),
        ),
        operator_args={
            "py_system_site_packages": False,
            "py_requirements": ["dbt-postgres"],
            "install_deps": True,
        },
    )

    start_task >> [tmp_venv_task_group, cached_venv_task_group] >> end_task


example_virtualenv()

Docker

The docker approach assumes users have a previously created Docker image, which should contain all the dbt pipelines and a profiles.yml, managed by the user.

The user has better environment isolation than when using local or virtualenv modes, but also more responsibility (ensuring the Docker container used has up-to-date files and managing secrets potentially in multiple places).

The other challenge with the docker approach is if the Airflow worker is already running in Docker, which sometimes can lead to challenges running Docker in Docker.

This approach can be significantly slower than virtualenv since it may have to build the Docker container, which is slower than creating a Virtualenv with dbt-core. If dbt is unavailable in the Airflow scheduler, the default LoadMode.DBT_LS will not work. In this scenario, users must use a parsing method that does not rely on dbt, such as LoadMode.MANIFEST.

Check the step-by-step guide on using the docker execution mode at Docker Execution Mode.

Example DAG:

docker_cosmos_dag = DbtDag(
    # ...
    execution_config=ExecutionConfig(
        execution_mode=ExecutionMode.DOCKER,
    ),
    operator_args={
        "image": "dbt-jaffle-shop:1.0.0",
        "network_mode": "bridge",
    },
)

Kubernetes

The kubernetes approach is a very isolated way of running dbt since the dbt run commands from within a Kubernetes Pod, usually in a separate host.

It assumes the user has a Kubernetes cluster. It also expects the user to ensure the Docker container has up-to-date dbt pipelines and profiles, potentially leading the user to declare secrets in two places (Airflow and Docker container).

The Kubernetes deployment may be slower than Docker and Virtualenv assuming that the container image is built (which is slower than creating a Python virtualenv and installing dbt-core) and the Airflow task needs to spin up a new Pod in Kubernetes.

Check the step-by-step guide on using the kubernetes execution mode at Kubernetes Execution Mode.

Example DAG:

    load_seeds = DbtSeedKubernetesOperator(
        task_id="load_seeds",
        project_dir=K8S_PROJECT_DIR,
        get_logs=True,
        schema="public",
        image=DBT_IMAGE,
        is_delete_operator_pod=False,
        secrets=[postgres_password_secret, postgres_host_secret],
        profile_config=ProfileConfig(
            profile_name="postgres_profile",
            target_name="dev",
            profile_mapping=PostgresUserPasswordProfileMapping(
                conn_id="postgres_default",
                profile_args={
                    "schema": "public",
                },
            ),
        ),
    )

AWS_EKS

The aws_eks approach is very similar to the kubernetes approach, but it is specifically designed to run on AWS EKS clusters. It uses the EKSPodOperator to run the dbt commands. You need to provide the cluster_name in your operator_args to connect to the AWS EKS cluster.

Example DAG:

postgres_password_secret = Secret(
    deploy_type="env",
    deploy_target="POSTGRES_PASSWORD",
    secret="postgres-secrets",
    key="password",
)

docker_cosmos_dag = DbtDag(
    # ...
    execution_config=ExecutionConfig(
        execution_mode=ExecutionMode.AWS_EKS,
    ),
    operator_args={
        "image": "dbt-jaffle-shop:1.0.0",
        "cluster_name": CLUSTER_NAME,
        "get_logs": True,
        "is_delete_operator_pod": False,
        "secrets": [postgres_password_secret],
    },
)

Azure Container Instance

Added in version 1.4.

Similar to the kubernetes approach, using Azure Container Instances as the execution mode gives a very isolated way of running dbt, since the dbt run itself is run within a container running in an Azure Container Instance.

This execution mode requires the user has an Azure environment that can be used to run Azure Container Groups in (see Azure Container Instance Execution Mode for more details on the exact requirements). Similarly to the Docker and Kubernetes execution modes, a Docker container should be available, containing the up-to-date dbt pipelines and profiles.

Each task will create a new container on Azure, giving full isolation. This, however, comes at the cost of speed, as this separation of tasks introduces some overhead. Please checkout the step-by-step guide for using Azure Container Instance as the execution mode

docker_cosmos_dag = DbtDag(
    # ...
    execution_config=ExecutionConfig(
        execution_mode=ExecutionMode.AZURE_CONTAINER_INSTANCE
    ),
    operator_args={
        "ci_conn_id": "aci",
        "registry_conn_id": "acr",
        "resource_group": "my-rg",
        "name": "my-aci-{{ ti.task_id.replace('.','-').replace('_','-') }}",
        "region": "West Europe",
        "image": "dbt-jaffle-shop:1.0.0",
    },
)

GCP Cloud Run Job

Added in version 1.7.

The gcp_cloud_run_job execution mode is particularly useful for users who prefer to run their dbt commands on Google Cloud infrastructure, taking advantage of Cloud Run’s scalability, isolation, and managed service capabilities.

For the gcp_cloud_run_job execution mode to work, a Cloud Run Job instance must first be created using a previously built Docker container. This container should include the latest dbt pipelines and profiles. You can find more details in the Cloud Run Job creation guide .

This execution mode allows users to run dbt core CLI commands in a Google Cloud Run Job instance. This mode leverages the CloudRunExecuteJobOperator from the Google Cloud Airflow provider to execute commands within a Cloud Run Job instance, where dbt is already installed. Similarly to the Docker and Kubernetes execution modes, a Docker container should be available, containing the up-to-date dbt pipelines and profiles.

Each task will create a new Cloud Run Job execution, giving full isolation. The separation of tasks adds extra overhead; however, that can be mitigated by using the concurrency parameter in DbtDag, which will result in parallelized execution of dbt models.

gcp_cloud_run_job_cosmos_dag = DbtDag(
    # ...
    execution_config=ExecutionConfig(execution_mode=ExecutionMode.GCP_CLOUD_RUN_JOB),
    operator_args={
        "project_id": "my-gcp-project-id",
        "region": "europe-west1",
        "job_name": "my-crj-{{ ti.task_id.replace('.','-').replace('_','-') }}",
    },
)

Airflow Async (experimental)

Added in version 1.7.0.

(Experimental) The airflow_async execution mode is a way to run the dbt resources from your dbt project using Apache Airflow’s Deferrable operators. This execution mode could be preferred when you’ve long running resources and you want to run them asynchronously by leveraging Airflow’s deferrable operators. With that, you would be able to potentially observe higher throughput of tasks as more dbt nodes will be run in parallel since they won’t be blocking Airflow’s worker slots.

In this mode, Cosmos adds a new operator, DbtCompileAirflowAsyncOperator, as a root task in the DbtDag or DbtTaskGroup. The task runs the dbt compile command on your dbt project which then outputs compiled SQLs in the project’s target directory. As part of the same task run, these compiled SQLs are then stored remotely to a remote path set using the remote_target_path: configuration. The remote path is then used by the subsequent tasks in the DAG to fetch (from the remote path) and run the compiled SQLs asynchronously using e.g. the DbtRunAirflowAsyncOperator. You may observe that the compile task takes a bit longer to run due to the latency of storing the compiled SQLs remotely (e.g. for the classic jaffle_shop dbt project, upon compiling it produces about 31 files measuring about 124KB in total, but on a local machine it took approximately 25 seconds for the task to compile & upload the compiled SQLs to the remote path)., however, it is still a win as it is one-time overhead and the subsequent tasks run asynchronously utilising the Airflow’s deferrable operators and supplying to them those compiled SQLs.

Note that currently, the airflow_async execution mode has the following limitations and is released as Experimental:

1. Airflow 2.8 or higher required: This mode relies on Airflow’s Object Storage feature, introduced in Airflow 2.8, to store and retrieve compiled SQLs.

2. Limited to dbt models: Only dbt resource type models are run asynchronously using Airflow deferrable operators. Other resource types are executed synchronously, similar to the local execution mode.

3. BigQuery support only: This mode only supports BigQuery as the target database. If a different target is specified, Cosmos will throw an error indicating the target database is unsupported in this mode.

4. ProfileMapping parameter required: You need to specify the ProfileMapping parameter in the ProfileConfig for your DAG. Refer to the example DAG below for details on setting this parameter.

5. Supports only full_refresh models: Currently, only full_refresh models are supported. To enable this, pass full_refresh=True in the operator_args of the DbtDag or DbtTaskGroup. Refer to the example DAG below for details on setting this parameter.

6. location parameter required: You must specify the location of the BigQuery dataset in the operator_args of the DbtDag or DbtTaskGroup. The example DAG below provides guidance on this.

7. No dataset emission: The async run operators do not currently emit datasets, meaning that Data-Aware Scheduling is not supported at this time. Future releases will address this limitation.

To start leveraging async execution mode that is currently supported for the BigQuery profile type targets you need to install Cosmos with the below additional dependencies:

astronomer-cosmos[dbt-bigquery, google]

Example DAG:

simple_dag_async = DbtDag(
    # dbt/cosmos-specific parameters
    project_config=ProjectConfig(
        DBT_ROOT_PATH / "original_jaffle_shop",
    ),
    profile_config=profile_config,
    execution_config=ExecutionConfig(
        execution_mode=ExecutionMode.AIRFLOW_ASYNC,
    ),
    render_config=RenderConfig(
        select=["path:models"],
        # test_behavior=TestBehavior.NONE
    ),
    # normal dag parameters
    schedule_interval=None,
    start_date=datetime(2023, 1, 1),
    catchup=False,
    dag_id="simple_dag_async",
    tags=["simple"],
    operator_args={"full_refresh": True, "location": "northamerica-northeast1"},
)

Known Issue:

The dag test command failed with the following error, likely because the trigger does not fully initialize during the dag test, leading to an uninitialized task instance. This causes the BigQuery trigger to attempt accessing parameters of the Task Instance that are not properly initialized.

[2024-10-01T18:19:09.726+0530] {base_events.py:1738} ERROR - unhandled exception during asyncio.run() shutdown
task: <Task finished name='Task-46' coro=<<async_generator_athrow without __name__>()> exception=AttributeError("'NoneType' object has no attribute 'dag_id'")>
Traceback (most recent call last):
  File "/Users/pankaj/Documents/astro_code/astronomer-cosmos/devenv/lib/python3.9/site-packages/airflow/providers/google/cloud/triggers/bigquery.py", line 138, in run
    yield TriggerEvent(
asyncio.exceptions.CancelledError

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/Users/pankaj/Documents/astro_code/astronomer-cosmos/devenv/lib/python3.9/site-packages/airflow/providers/google/cloud/triggers/bigquery.py", line 157, in run
    if self.job_id and self.cancel_on_kill and self.safe_to_cancel():
  File "/Users/pankaj/Documents/astro_code/astronomer-cosmos/devenv/lib/python3.9/site-packages/airflow/providers/google/cloud/triggers/bigquery.py", line 126, in safe_to_cancel
    task_instance = self.get_task_instance()  # type: ignore[call-arg]
  File "/Users/pankaj/Documents/astro_code/astronomer-cosmos/devenv/lib/python3.9/site-packages/airflow/utils/session.py", line 97, in wrapper
    return func(*args, session=session, **kwargs)
  File "/Users/pankaj/Documents/astro_code/astronomer-cosmos/devenv/lib/python3.9/site-packages/airflow/providers/google/cloud/triggers/bigquery.py", line 102, in get_task_instance
    TaskInstance.dag_id == self.task_instance.dag_id,
AttributeError: 'NoneType' object has no attribute 'dag_id'

Invocation Modes

Added in version 1.4.

For ExecutionMode.LOCAL execution mode, Cosmos supports two invocation modes for running dbt:

1. InvocationMode.SUBPROCESS: In this mode, Cosmos runs dbt cli commands using the Python subprocess module and parses the output to capture logs and to raise exceptions.

2. InvocationMode.DBT_RUNNER: In this mode, Cosmos uses the dbtRunner available for dbt programmatic invocations to run dbt commands. In order to use this mode, dbt must be installed in the same local environment. This mode does not have the overhead of spawning new subprocesses or parsing the output of dbt commands and is faster than InvocationMode.SUBPROCESS. This mode requires dbt version 1.5.0 or higher. It is up to the user to resolve Airflow and dbt dependencies conflicts when using this mode.

The invocation mode can be set in the ExecutionConfig as shown below:

from cosmos.constants import InvocationMode

dag = DbtDag(
    # ...
    execution_config=ExecutionConfig(
        execution_mode=ExecutionMode.LOCAL,
        invocation_mode=InvocationMode.DBT_RUNNER,
    ),
)

If the invocation mode is not set, Cosmos will attempt to use InvocationMode.DBT_RUNNER if dbt is installed in the same environment as the worker, otherwise it will fall back to InvocationMode.SUBPROCESS.