SDK, CLI, and Boto3 deployments

SageMaker Python SDK

The notebook-native layer — ML concepts in, API mechanics hidden:

from sagemaker.model import Model

model = Model(image_uri=image, model_data=s3_artifact, role=role)
predictor = model.deploy(
    initial_instance_count=1,
    instance_type="ml.m5.xlarge",
)
predictor.predict(payload)

One deploy() = three objects created and a Predictor for calling the result. Perfect for experimentation; in pipelines its implicitness can hide what will exist afterward.

Boto3

The explicit layer — every object spelled out, which is why Lambda and pipeline code uses it:

import boto3
sm = boto3.client("sagemaker")

sm.create_model(ModelName=..., PrimaryContainer=..., ExecutionRoleArn=...)
sm.create_endpoint_config(EndpointConfigName=..., ProductionVariants=[...])
sm.create_endpoint(EndpointName=..., EndpointConfigName=...)

runtime = boto3.client("sagemaker-runtime")
runtime.invoke_endpoint(EndpointName=..., Body=payload)

Note the two clients: sagemaker builds, sagemaker-runtime invokes.

AWS CLI

The same APIs from a shell — deploy scripts, cron, quick checks:

aws sagemaker create-endpoint \
  --endpoint-name churn-prod \
  --endpoint-config-name churn-config-v7

aws sagemaker-runtime invoke-endpoint \
  --endpoint-name churn-prod \
  --body fileb://payload.json out.json

Same control/data-plane split: aws sagemaker vs aws sagemaker-runtime.

Three ladders to the same endpoint

1. Trained model artifact in S3

   A model.tar.gz plus an inference image URI — the raw material every ladder starts from.

2. Who is deploying?

   The surfaces differ by altitude, not capability: a person exploring in a notebook, automation inside Lambda or a pipeline step, or a shell script — all drive the same SageMaker control plane.

3. SDK: model.deploy()

   The SageMaker Python SDK wraps the whole chain: one call creates the model, the endpoint config, and the endpoint — and hands back a Predictor for calling it.

4. Boto3: create_model

   Rung one of the explicit ladder: name the container image and artifact, attach the execution role.

5. create_endpoint_config

   Rung two: production variants — instance type, count, weights. The config is the versioned object a later UpdateEndpoint swaps in.

6. create_endpoint

   Rung three: stand the endpoint up from the config. It fails if the name already exists — the idempotency trap the callout below covers.

7. CLI: aws sagemaker …

   The same three calls as shell commands — create-model, create-endpoint-config, create-endpoint. Invocation moves to the sibling aws sagemaker-runtime command.

8. The same running endpoint

   Whichever ladder you climb, describe_endpoint shows identical resources: a model, a config, an InService endpoint — invoked through the sagemaker-runtime client, never the sagemaker one.