Ksama Arora

(I) DEPLOY MLFLOW MODEL TO A MANAGED ONLINE ENDPOINT

Note: When you deploy MLFlow models to an online endpoint, a scoring script and environment are automatically generated, and no need to create them.

To deploy Mlflow model, you need

• Model files stored on a local path OR with registered models
• Compute configuration for deployment ○ instance_type: VM size to use ○ instance_count: Number of instances to use

Code to deploy a MLflow model to managed online endpoint:

from azure.ai.ml.entities import Model, ManagedOnlineDeployment
from azure.ai.ml.constants import AssetTypes

# create a blue deployment
model = Model(
    path="./model",
    type=AssetTypes.MLFLOW_MODEL,
    description="my sample mlflow model",
)

blue_deployment = ManagedOnlineDeployment(
    name="blue",
    endpoint_name="endpoint-example",
    model=model,
    instance_type="Standard_F4s_v2",
    instance_count=1,
)

ml_client.online_deployments.begin_create_or_update(blue_deployment).result()

Deploy to a specific model among mutiple models: (IMP)

Since only one model deployed, 100% of traffic is to this model. However if multiple models are deployed to same endpoint, and to direct traffic to a specific deployment use this code

# blue deployment takes 100 traffic
endpoint.traffic = {"blue": 100}
ml_client.begin_create_or_update(endpoint).result()

IMP NOTE: 25% traffic goes to blue model and 75% goes to green model. Code snippet implementing desired blue/green configuration:

endpoint.traffic = {"blue": 25, "green": 75} 
ml_client.begin_create_or_update(endpoint).result()

To delete endpoint and all associated deployments

ml_client.online_endpoints.begin_delete(name="endpoint-example")