Skip to main content
Version: 0.93

snorkelflow.sdk.MLflowDeployment

class snorkelflow.sdk.MLflowDeployment(deployment_uid, **kwargs)

Bases: Deployment

MLflowDeployment class represents a deployment of a Snorkel Flow application as an MLflow model. This class inherits from the Deployment class and provides additional functionality specific to MLflow deployments. Since MLflow is the only supported deployment type at the moment, this class is the only implementation of the Deployment class. This class is intended to be an interface between Snorkel Flow and MLflow.

A typical workflow for deploying an application to production is as follows:

from snorkelflow.sdk import MLflowDeployment

deployment = MLflowDeployment.create("my-application", "my-deployment")
deployment.execute(df)
deployment.download('/path/to/download')

Follow with MLflow client calls to deploy your workflow to production.

If you want to get a deployment by its name, you can use the following code:

from snorkelflow.sdk import MLflowDeployment

deployments = MLflowDeployment.list(application="my-application")
matched_deployment = next((deployment for deployment in deployments if deployment.name == "my-deployment"), None)
if matched_deployment is None:
    print("Deployment not found")
__init__(deployment_uid, **kwargs)

Initializes a Deployment object. This constructor should not be called directly. Instead, use the Deployment.create method to create a deployment.

Parameters:

deployment_uid (int) – The unique identifier of the deployment.

Methods

__init__(deployment_uid, **kwargs)

Initializes a Deployment object.

create(application[, name, registry_type, ...])

Create a new MLflow deployment from an application.

delete(deployment_uid)

Delete a deployment based on the provided identifier.

download(local_path[, download_as_zip, ...])

Download an MLflow deployment to the specified local path.

execute(df)

Execute a deployment on a pandas DataFrame.

get(deployment_uid)

Fetches a Deployment object based on the provided identifier.

list([application])

List all deployments for an application.

register(tracking_server_uri, *[, ...])

Register an MLflowModel Deployment with an external MLflow model registry.

update(name)

Update the metadata for the deployment.

classmethod create(application, name=None, registry_type='INTERNAL', model_registry_uid=None, signature=True, experiment_name=None)

Create a new MLflow deployment from an application.

Examples

>>> from snorkelflow.sdk import MLflowDeployment
>>> my_deployment = MLflowDeployment.create(application="my-app", name="my-deployment")
{'deployment_uid': <deployment_uid>}
>>> my_deployment.name
"my-deployment"
Parameters:

  • application (Union[str, int]) – The name or UID of the application to be exported

  • name (Optional[str], default: None) – The name of the deployment. If this is not provided, deployment name will be created using the application name. Deployment names are not unique and re-running this method with the same name will create a new deployment. To register a model to Databricks Unity Catalog, the name should be in the form of “<catalog>.<schema>.<model>”.

  • registry_type (str, default: 'INTERNAL') – The type of the model registry, “INTERNAL” or “EXTERNAL”.

  • model_registry_uid (Optional[int], default: None) – The UID of the external model registry. If registry_type is “INTERNAL”, model_registry_uid is not required.

  • signature (bool, default: True) – Whether to add a signature for the input_schema in the MLmodel yaml file. If True, a signature will be added. If False, no signature will be added.

  • experiment_name (Optional[str], default: None) – The name of the experiment to register the model under, used the deployment name if not provided. This is required for certain model registries such as Databricks, where this must an absolute path, e.g. “/Users/<username>/<experiment_name>”.

Returns:

MLflowDeployment object containing information about the deployed application

Return type:

MLflowDeployment

download(local_path, download_as_zip=False, overwrite=False)

Download an MLflow deployment to the specified local path. To see how to use this downloaded package, please go to https://mlflow.org/.

Examples

>>> deployment.download(local_path='./my-folder', download_as_zip=True)
Successfully downloaded deployment to ./my-folder/my-deployment-name.zip
>>> deployment.download(local_path='./my-folder', download_as_zip=False)
Successfully downloaded deployment to ./my-folder/my-deployment-name
Parameters:

  • local_path (str) – The local path to download the deployment to.

  • download_as_zip (bool, default: False) – Whether to download the deployment as a zip file. If False, the deployment will be downloaded as a directory.

  • overwrite (bool, default: False) – Whether to overwrite the local path if it already exists. If False, an error will be thrown if the path already exists.

Return type:

None

register(tracking_server_uri, *, experiment_name=None, model_name=None, mlflow_s3_endpoint_url=None, aws_access_key_id=None, aws_secret_access_key=None, basic_auth_username=None, basic_auth_password=None, bearer_auth_token=None, timeout_seconds=600)

Register an MLflowModel Deployment with an external MLflow model registry.

Examples

>>> from snorkelflow.sdk import MLflowDeployment
>>> deployment = MLflowDeployment.create(application="my-app")
>>> deployment.register(tracking_server_uri="http://localhost:5000")
<Response [200]>
Parameters:

  • tracking_server_uri (str) – URI of the tracking server to export the deployment to

  • experiment_name (Optional[str], default: None) – The name of the experiment to register the model under, used default if not provided

  • model_name (Optional[str], default: None) – The name of the model to register the deployment under. Defaults to the parent application name

  • mlflow_s3_endpoint_url (Optional[str], default: None) – The endpoint of the S3-compatible storage service used by the tracking server

  • aws_access_key_id (Optional[str], default: None) – The access key for the s3-compatible storage service used by the tracking server

  • aws_secret_access_key (Optional[str], default: None) – The secret key for the s3-compatible storage service used by the tracking server

  • basic_auth_username (Optional[str], default: None) – Username needed for access to servers with basic authentication enabled

  • basic_auth_password (Optional[str], default: None) – Password needed for access to servers with basic authentication enabled

  • bearer_auth_token (Optional[str], default: None) – Bearer token needed for access to servers with bearer authentication enabled

  • timeout_seconds (int, default: 600) – Maximum amount of time to wait for the job to complete before giving up

Returns:

A dictionary describing the model that has been exported to the external service. Resembles a mlflow.model.models.ModelInfo object with extra information on the active Experiment.

Return type:

Dict