IMPORTANT NOTE: This package is no longer maintained and not adapted for mlflow>2.*.

Openshift Mlflow Deployment Plugin

Mlflow deployment plugin for openshift. This plugin offers the possibility to deploy mlflow packed models into openshift using the regular mlflow deployment command line interface.

Installation

1. Install the mlflow openshift plugin: pip install mlflow-openshift.
2. Make sure the openshift CLI tool is installed by calling oc in the command line. If not, you can find an installation tutorial here

Get Started

1. Get your login token from the openshift web-ui and use it to log in. You can find it on the top right > question mark > about > command line tools.

   oc login <token>
   

2. Navigate to the openshift project you want to the deploy the model. Make sure you have admin priviliges in that project.

   oc project <my-project>
   

   You can validate if your current user has admin rights for the project by executing this command:

   oc get rolebindings admin -n <my-project>
   

3. Setup the mlflow (and s3/minio) environment variables:

   AWS_ACCESS_KEY_ID=<>
   AWS_SECRET_ACCESS_KEY=<>
   MLFLOW_S3_ENDPOINT_URL=<>
   

Create a Deployment

Creates all necessary artifacts for a model deployment in openshift, i.e. hosting the model in the specified container image and putting and nginx basic authentication proxy in front of the container to publisch an https endpoint.

The succesful deployment will return the created https host. Requests can be sent against mlflow's default /invocations endpoint.

Mandatory config items

--name
--model-uri
--docker-registry
--image
--tag
--auth_user
--auth_password

Optional config items:

--cpu_limit -> default: `1`
--cpu_request -> default: `100m`
--mem_limit -> default: `512Mi`
--mem_request -> default: `256Mi`
--gunicorn_workers -> default: `1`

Example: MLflow CLI

mlflow deployments create -t openshift \
    --name <name> \
    --model-uri <model-uri>
    --config docker_registry=<docker-registry> \
    --config image=<image> \
    --config tag=<tag>

Example: python mlflow API

from mlflow.deployments import get_deploy_client
target_uri = 'openshift'
openshift_client = get_deploy_client(target_uri)

openshift_client.create_deployment(
    <name>,
    <model-uri>,
        "docker_registry": <docker_registry>,
        "image": <image>,
        "tag": <tag>
    }
)

Updating an existing Deployment

Updates an existing model deployment in openshift. It can either update the model_uri and/or the config items describing the container image (all three of them need to be provided), i.e image, docker_registry, tag.

Example: MLflow CLI

mlflow deployments update -t openshift \
    --name <name> \
    --model-uri <model-uri>

Example: python mlflow API

from mlflow.deployments import get_deploy_client
target_uri = 'openshift'
openshift_client = get_deploy_client(target_uri)

openshift_client.update_deployment(
    <name>,
    model_uri=<model-uri>
)

Deleting a Deyployment

Deletes the deployment and resources (openshift artifacts like routes).

Example: MLflow CLI

mlflow deployments delete -t openshift --name <name>

Example: python mlflow API

from mlflow.deployments import get_deploy_client
target_uri = 'openshift'
openshift_client = get_deploy_client(target_uri)

openshift_client.delete_deployment(<name>)

Listing all Mlflow Deplyoments

Lists all mlflow deployments in the current openshift project.

Example: MLflow CLI

mlflow deployments list -t openshift

Example: python mlflow API

from mlflow.deployments import get_deploy_client
target_uri = 'openshift'
openshift_client = get_deploy_client(target_uri)

openshift_client.list_deployments()

Get Deplyoment Information

Retrieves raw, detailed information for the deployment.

Example: MLflow CLI

mlflow deployments get -t openshift --name <name>

Example: python mlflow API

from mlflow.deployments import get_deploy_client
target_uri = 'openshift'
openshift_client = get_deploy_client(target_uri)

openshift_client.get_deployment(<name>)

Predict with Deployment

Makes predictions using the specified deployment name. This can be used for making batch predictions using the openshift infrastrucutre, e.g. in automated daily/weekly pipelines. This option is only available for the python mlflow API. However, the REST endpoint can of course be called by and REST capable service.

Example: python mlflow API

For a iris flower dataset model

from mlflow.deployments import get_deploy_client
target_uri = 'openshift'
openshift_client = get_deploy_client(target_uri)

df = pd.DataFrame(
    columns=["sepalLength", "sepalWidth", "petalWidth"],
    data=[[0, 1, 0], [0, 1, 1]]
)

predictions = openshift_client.predict(<name>, df)