Correct:
* Create an environment.
Incorrect:
* Delete the Python 3.6 - AzureML kernel.
* Delete the Python 3.8 - AzureML kernel.
Note:
Before you can add a new Jupyter kernel on an Azure Machine Learning compute instance terminal, you must create a Conda environment.
Required Workflow
To officially provision and expose the new kernel to your Azure Machine Learning studio Notebooks, you need to execute the following full process from your terminal session:
Create the environment: Provision a new isolated environment (e.g., using conda create -n newenv python=3.10).
Activate the environment: Run conda activate newenv.
Install dependencies: Add the required ipykernel package using conda install ipykernel or pip install ipykernel.
Register the kernel: Bind the new environment configuration to the global Jupyter directory by running:
python -m ipykernel install --user --name newenv --display-name "My New Kernel" Reference:
https://docs.azure.cn/en-us/machine-learning/how-to-access-terminal

The best strategy for controlled testing of a new model version in Azure Machine Learning is Blue-Green Deployment, often referred to as a safe rollout.
This approach allows you to deploy a new model version alongside the current one within the same Managed Online Endpoint without disrupting existing users.
Key Features of Blue-Green Deployment in Azure ML
Simultaneous Versions: Both the current "Blue" and new "Green" models run concurrently on the same endpoint.
*-> Traffic Shifting: You can use the endpoint's load balancer to allocate a specific percentage (e.g., 10%) of live production traffic to the new version.
Mirrored Traffic: For even lower risk, you can test the new model with mirrored traffic, where production requests are copied to the new model for validation without using its responses for the end user.
Instant Rollback: If the new model performs poorly, you can instantly shift 100% of traffic back to the original version.
Deployment Headers: You can bypass general traffic splitting to test the "Green" deployment specifically by adding an azureml-model-deployment header to your HTTP requests.
Reference:
https://learn.microsoft.com/en-us/azure/machine-learning/how-to-safely-rollout-online-endpoints

Correct:
* In Microsoft Foundry, turn on Tracing for the prompt flow of the project and execute test runs to produce trace data.
Incorrect:
* Create prompt variants and compare their outputs in the Evaluation experience.
* Use the prompt flow SDK to enable tracing for the flow before executing runs. Then run the flow to generate traceable results.
Note:
In Azure AI Foundry, you can capture and compare these metrics by enabling Tracing and using the Bulk Test feature. This allows you to systematically evaluate different prompt variants against a common dataset.
Steps to Evaluate and Compare Prompt Variants
*-> 1. Enable Tracing
Navigate to your Prompt Flow project.
Locate the Tracing toggle at the top of the flow authoring page.
Switch it to On.
This ensures every execution captures latency, token counts, and node-level inputs/outputs.
2. Create Prompt Variants
Within your flow, identify the LLM node you want to test.
Click Variants to create multiple versions of your prompt (e.g., Variant_0, Variant_1).
This allows you to test different instructions or few-shot examples side-by-side.
3. Run a Bulk Test (Evaluation)
4. Analyze the Results
Reference:
https://www.linkedin.com/pulse/streamlining-generative-ai-development-azure-foundry-tracing-taneja-mbwze

Using a single pipeline triggered by both a schedule and data drift alerts ensures consistent retraining logic and avoids duplication. This approach minimizes operational overhead and maintenance complexity. Creating multiple pipelines can lead to inconsistencies, duplicated code, and increased effort when updating retraining logic or dependencies.