How can I deploy a re-trained Sagemaker model to an endpoint?

Question

With an sagemaker.estimator.Estimator , I want to re- deploy a model after retraining (calling fit with new data).

When I call this

estimator.deploy(initial_instance_count=1, instance_type='ml.m5.xlarge')

I get an error

botocore.exceptions.ClientError: An error occurred (ValidationException) 
when calling the CreateEndpoint operation: 
Cannot create already existing endpoint "arn:aws:sagemaker:eu-east- 
1:1776401913911:endpoint/zyx".

Apparently I want to use functionality like UpdateEndpoint . How do I access that functionality from this API?

Answer 1

update_endpoint has been deprecated since AFAIK. To re-create the UpdateEndpoint functionality from this API itself and deploy a newly fit training job to an existing endpoint, we could do something like this (this example uses the sagemaker sklearn API):

from sagemaker.sklearn.estimator import SKLearn

sklearn_estimator = SKLearn(
    entry_point=model.py,
    instance_type=<instance_type>,
    framework_version=<framework_version>,
    role=<role>,
    dependencies=[
         <comma seperated names of files>
    ],
    hyperparameters={
         'key_1':value,
         'key_2':value,
         ...
    }
)

sklearn_estimator.fit()

sm_client = boto3.client('sagemaker')

# Create the model
sklearn_model = sklearn_estimator.create_model()

# Define an endpoint config and an endpoint
endpoint_config_name = 'endpoint-' + datetime.utcnow().strftime("%Y%m%d%H%m%s")
current_endpoint = endpoint_config_name

# From the Model : create the endpoint config and the endpoint
sklearn_model.deploy(
    initial_instance_count=<count>,
    instance_type=<instance_type>,
    endpoint_name=current_endpoint
)

# Update the existing endpoint if it exists or create a new one
try:
    sm_client.update_endpoint(
        EndpointName=DESIRED_ENDPOINT_NAME, # The Prod/Existing Endpoint Name
        EndpointConfigName=endpoint_config_name
    )
except Exception as e:
    try:
        sm_client.create_endpoint(
            EndpointName=DESIRED_ENDPOINT_NAME, # The Prod Endpoint name
            EndpointConfigName=endpoint_config_name
        )
    except Exception as e:
        logger.info(e)

sm_client.delete_endpoint(EndpointName=current_endpoint)

Answer 2

Yes, under the hood the model.deploy creates a model, an endpoint configuration and an endpoint. When you call again the method from an already-deployed, trained estimator it will create an error because a similarly-configured endpoint is already deployed. What I encourage you to try:

• use the update_endpoint=True parameter. From the SageMaker SDK doc : "Additionally, it is possible to deploy a different endpoint configuration, which links to your model, to an already existing SageMaker endpoint. This can be done by specifying the existing endpoint name for the endpoint_name parameter along with the update_endpoint parameter as True within your deploy() call."

• Alternatively, if you want to create a separate model you can specify a new model_name in your deploy