Submit flows directly to dynamic infrastructure

Prefect allows you to submit workflows directly to different infrastructure types without requiring a deployment. This enables you to dynamically choose where your workflows run based on their requirements, such as:

• Training machine learning models that require GPUs
• Processing large datasets that need significant memory
• Running lightweight tasks that can use minimal resources

​

Benefits

Submitting workflows directly to dynamic infrastructure provides several advantages:

• Dynamic resource allocation: Choose infrastructure based on workflow requirements at runtime
• Cost efficiency: Use expensive infrastructure only when needed
• Consistency: Ensure workflows always run on the appropriate infrastructure type
• Simplified workflow management: No need to create and maintain deployments for different infrastructure types

​

Supported infrastructure

Direct submission of workflows is currently supported for the following infrastructures:

Each package can be installed using pip, for example:

pip install prefect-docker

​

Prerequisites

Before submitting workflows to specific infrastructure, you’ll need:

1. A work pool for each infrastructure type you want to use
2. Object storage to associate with your work pool(s)
3. A storage block to use for result storage

​

Setting up work pools and storage

​

Creating a work pool

Create work pools for each infrastructure type using the Prefect CLI:

prefect work-pool create NAME --type WORK_POOL_TYPE

For detailed information on creating and configuring work pools, refer to the work pools documentation.

​

Configuring work pool storage

To enable Prefect to run workflows in remote infrastructure, work pools need an associated storage location to store serialized versions of submitted workflows.

Configure storage for your work pools using one of the supported storage types:

prefect work-pool storage configure s3 WORK_POOL_NAME \
    --bucket BUCKET_NAME \
    --aws-credentials-block-name BLOCK_NAME

To allow Prefect to upload and download serialized workflows, you can create a block containing credentials with permission to access your configured storage location.

If a credentials block is not provided, Prefect will use the default credentials (e.g., a local profile or an IAM role) as determined by the corresponding cloud provider.

You can inspect your storage configuration using:

prefect work-pool storage inspect WORK_POOL_NAME

​

Configuring result storage

To retrieve the return value from an infrastructure-bound workflow, you’ll need to configure result storage for your workflow.

The result storage location must be accessible by the workflow submitter and the remote infrastructure, so most of the time you’ll want to use an object storage location like S3, GCS, or Azure Blob storage.

Return values for submitted workflows must be serializable for storage and retrieval.

from prefect import flow
from prefect.filesystems import LocalFileSystem
from prefect_docker.experimental import docker


result_storage = LocalFileSystem(basepath="/tmp/results")
result_storage.save("result-storage", overwrite=True)


@docker(
    work_pool="above-ground",
    volumes=["/tmp/results:/tmp/results"],
)
@flow(result_storage=result_storage)
def run_in_docker(name: str):
    return(f"Hello, {name}!")


print(run_in_docker("world")) # prints "Hello, world!"

​

Submitting workflows to specific infrastructure

To submit a flow to specific infrastructure, use the appropriate decorator for that infrastructure type.

Here’s an example using @kubernetes:

from prefect import flow
from prefect_kubernetes.experimental.decorators import kubernetes


# Submit `my_remote_flow` to run in a Kubernetes job
@kubernetes(work_pool="olympic")
@flow
def my_remote_flow(name: str):
    print(f"Hello {name}!")

@flow
def my_flow():
    my_remote_flow("Marvin")

# Run the flow
my_flow()

When you run this code on your machine, my_flow will execute locally, while my_remote_flow will be submitted to run in a Kubernetes job.

​

Customizing infrastructure configuration

You can override the default configuration by providing additional kwargs to the infrastructure decorator:

from prefect import flow
from prefect_kubernetes.experimental.decorators import kubernetes


@kubernetes(
    work_pool="my-kubernetes-pool",
    namespace="custom-namespace"
)
@flow
def custom_namespace_flow():
    pass

Any kwargs passed to the infrastructure decorator will override the corresponding default value in the base job template for the specified work pool.