Model Versioning

Overview

Managing model versions with SageMaker Model Registry.

Model Registry

Centralized repository for model versioning.

Key Concepts

Concept	Description
Model Package Group	Collection of model versions
Model Package	Single model version
Approval Status	Pending, Approved, Rejected
Model Metrics	Performance metrics

Creating a Model Package Group

import boto3

sm_client = boto3.client("sagemaker")

sm_client.create_model_package_group(
    ModelPackageGroupName="my-model-group",
    ModelPackageGroupDescription="Production models for customer churn"
)

Registering a Model

from sagemaker.model import Model

model = Model(
    image_uri=inference_image,
    model_data="s3://bucket/model.tar.gz",
    role=role
)

model_package = model.register(
    model_package_group_name="my-model-group",
    content_types=["application/json"],
    response_types=["application/json"],
    inference_instances=["ml.m5.large"],
    transform_instances=["ml.m5.large"],
    approval_status="PendingManualApproval",
    model_metrics={
        "ModelQuality": {
            "Statistics": {
                "ContentType": "application/json",
                "S3Uri": "s3://bucket/metrics.json"
            }
        }
    }
)

Approval Workflow

# Approve a model
sm_client.update_model_package(
    ModelPackageArn=model_package_arn,
    ModelApprovalStatus="Approved"
)

# Reject a model
sm_client.update_model_package(
    ModelPackageArn=model_package_arn,
    ModelApprovalStatus="Rejected"
)

Model Lineage

Track the complete lifecycle of a model.

from sagemaker.lineage.context import Context
from sagemaker.lineage.artifact import Artifact

# Query lineage
artifacts = Artifact.list(source_uri="s3://bucket/model.tar.gz")

SageMaker Experiments

Track experiments and compare results.

from sagemaker.experiments import Run

with Run(experiment_name="my-experiment", run_name="run-1") as run:
    run.log_parameter("learning_rate", 0.01)
    run.log_metric("accuracy", 0.95)
    run.log_artifact(name="model", value="s3://bucket/model.tar.gz")

Best Practices

!!! tip "Versioning Best Practices" 1. Use Model Registry for all production models 2. Implement approval workflows for governance 3. Track metrics with each model version 4. Use lineage for reproducibility 5. Tag models with relevant metadata

Exam Tips

!!! warning "Key Points" - Model Registry for versioning and governance - Approval status gates deployment - Lineage tracks data→training→model→deployment - Experiments for comparing model runs