Drag and Drop Question You have an Azure Machine Learning workspace named WS1. You plan to use WS1 to train two models named model1 and model2. For model1, you plan to use automated machine learning.

Question

Exhibit

Answer Area

Explanation

This question tests the ability to select appropriate Azure Machine Learning compute targets based on specific operational requirements for training different machine learning models.

Approach. To answer correctly, we need to match the requirements for each model to the capabilities of the available compute targets:

1. For model1 (Automated ML, supports auto-shutdown/auto-start based on a schedule):

   • Azure Machine Learning compute instance is the correct choice. A compute instance is a single-node, managed development workstation in the cloud. It is designed for development and experimentation and perfectly supports automated machine learning. Crucially, it can be configured with idle shutdown policies and scheduled start/stop times, directly meeting the auto-shutdown/auto-start requirement for cost management when not in active use.

2. For model2 (Azure Machine Learning designer, supports low-priority Azure Virtual Machines):

   • Azure Machine Learning Kubernetes is the correct choice. An Azure Machine Learning Kubernetes compute target allows you to attach an existing Azure Kubernetes Service (AKS) cluster to your workspace. AKS clusters can be configured with node pools that use Azure Spot Virtual Machines (often referred to as low-priority VMs). These spot instances provide a cost-effective solution for running scalable training jobs, including those defined and executed via Azure Machine Learning designer pipelines, as long as the workloads are fault-tolerant. While Azure Machine Learning compute clusters also support low-priority VMs, Kubernetes offers broader flexibility and is a valid target for Designer pipelines that can leverage spot instances for cost optimization.

Common mistakes.

• common_mistake. 1. Using Azure Machine Learning compute cluster for model1: While a compute cluster can run automated ML, its primary auto-management feature is auto-scaling based on the training workload (scaling nodes up/down). It does not offer direct auto-shutdown/auto-start of the entire cluster on a fixed schedule in the same manner as a compute instance, which is designed for scheduled cost optimization for interactive or intermittent use.

1. Using Azure Machine Learning compute instance for model2: A compute instance is a single-node VM and is not typically used for scalable training with 'low-priority Azure Virtual Machines' in the context of a cluster for cost optimization. While it's technically an Azure VM, the requirement for 'low-priority VMs' usually implies leveraging a cluster's ability to use spot instances for distributed or larger-scale jobs.
2. Using Azure Machine Learning Kubernetes for model1: Kubernetes is more complex and typically used for deployment or more advanced training scenarios. For a simple auto-shutdown/auto-start requirement for an Automated ML run, a compute instance is a far more straightforward and cost-effective solution.
3. Using Azure Machine Learning compute cluster for model2 (if Kubernetes is the intended answer): Azure Machine Learning compute clusters do support low-priority VMs and are a very common target for Azure ML Designer. However, given the need to select distinct compute targets and the explicit solution provided (Kubernetes for model2), the question likely aims to differentiate between the capabilities and common use cases. Both can provide low-priority VMs for designer jobs, but Kubernetes offers more control over the underlying infrastructure and might be preferred in scenarios requiring specific container orchestration or existing AKS integration.

Concept tested. Understanding the capabilities, use cases, and cost-management features of different Azure Machine Learning compute targets, specifically Compute Instances, Compute Clusters, and Kubernetes (AKS), and matching them to specific model training methodologies (Automated ML, Designer) and operational requirements (auto-shutdown/start, low-priority VMs).

Reference. null

No community discussion yet for this question.