fix: MLFlow E2E Example Notebook (5513)

[aviruthen]

Description

The existing MLflow E2E notebook at v3-examples/ml-ops-examples/v3-mlflow-train-inference-e2e-example.ipynb has several issues: (1) Uses registered_model.latest_versions which was removed in MLflow 3.x - should use client.search_model_versions() instead; (2) Uses Session.boto_region_name as a class attribute without instantiation - should be Session().boto_region_name; (3) Step 7 uses raw boto3 sagemaker-runtime client for inference instead of the V3 core_endpoint.invoke() pattern used in other V3 example notebooks; (4) The notebook references mlflow==3.4.0 but uses deprecated MLflow 2.x API patterns incompatible with MLflow 3.x.

Related Issue

Related issue: 5513

Changes Made

• v3-examples/ml-ops-examples/v3-mlflow-train-inference-e2e-example.ipynb
• v3-examples/inference-examples/train-inference-e2e-example.ipynb

AI-Generated PR

This PR was automatically generated by the PySDK Issue Agent.

• Confidence score: 85%
• Classification: bug
• SDK version target: V3

Merge Checklist

• Changes are backward compatible
• Commit message follows prefix: description format
• Unit tests added/updated
• Integration tests added (if applicable)
• Documentation updated (if applicable)

[mufaddal-rohawala]

🤖 AI Code Review

This PR fixes several real issues in example notebooks: MLflow 3.x API compatibility, incorrect Session class attribute access, and replacing raw boto3 calls with V3-native core_endpoint.invoke(). The changes are well-motivated and align with V3 SDK conventions. A few minor concerns around the Session() instantiation pattern and the core_endpoint.invoke() API usage.

[mufaddal-rohawala]

While Session().boto_region_name is correct (it needs instantiation), this creates a throwaway Session object just to get the region. Consider using boto3.session.Session().region_name for a lighter-weight approach, or storing the Session() instance for reuse later in the notebook:

sagemaker_session = Session()
AWS_REGION = sagemaker_session.boto_region_name

This avoids creating multiple Session objects if it's used elsewhere in the notebook.

[mufaddal-rohawala]

Same comment as the other notebook — consider storing the Session() instance in a variable for reuse rather than creating a throwaway object:

sagemaker_session = Session()
AWS_REGION = sagemaker_session.boto_region_name

[mufaddal-rohawala]

Good change replacing raw boto3 sagemaker-runtime client with core_endpoint.invoke() — this aligns with the V3 tenet that all subpackages should use sagemaker-core rather than calling boto3 directly.

However, please verify the exact parameter names for core_endpoint.invoke(). The sagemaker-core Endpoint.invoke() method may use Body and ContentType (PascalCase matching the API model) rather than body and content_type (snake_case). Similarly, the response attribute might be Body rather than body. If the snake_case versions work, that's fine — just want to make sure this has been tested.

[mufaddal-rohawala]

Minor: The order_by parameter value 'version_number DESC' — please verify this is the correct field name for MLflow's search_model_versions. The MLflow documentation uses 'version_number DESC' in some examples but older versions may use different field names. Since this notebook targets mlflow==3.4.0, it would be good to confirm this works with that specific version.

[mufaddal-rohawala]

Good defensive check raising ValueError with a descriptive message including the model name. This follows the SDK convention of fail-fast validation with specific exceptions.

[mufaddal-rohawala]

The comment says 'dependencies' parameter is deprecated and suggests configure_for_torchserve() — is this accurate? If so, should the example code itself be updated to use the non-deprecated approach rather than just adding a comment about it? Leaving deprecated code in an example notebook may confuse users.