Path8.Mod1.c - Intro to DevOps Principles for ML - Compute Targets

Question 1

These are all the possible Compute Targets for Training:
- Two On-Prem
- Four AML-Dedicated
- Four Azure Data Services
- One Open Source

Answer 1

On-Prem:
* Local Personal Computer
* Remote VMs

AML-Dedicated:
* AML Compute Clusters
* AML Serverless Compute
* AML Compute Instance
* AML Kubernetes

Azure Data Services:
* Azure Databricks
* Azure Data Lake Analytics
* Azure HDInsights
* Azure Batch

Open Source:
* Apache Spark pools (preview)

Question 2

Machine Learning Pipelines can use any Compute Target except for this one

Answer 2

Your Local Computer LOL~

Question 3

Automated Machine Learning (AutoML) can use every Compute Target except these four:
- One AML-Dedicated
- Three Azure Data Services

Answer 3

• AML Kubernetes
• Azure Data Lake Analytics
• Azure HDInsight
• Azure Batch

Surprisingly, AutoML can use Azure Databricks for a Compute Target…

Question 4

CI ApSpP AzDb

Automated Machine Learning (AutoML) can use these three Compute Targets with certain limitations; know the limitation

Answer 4

• Compute Instances, only through the SDK
• Apache Spark Pools (preview), only through SDK local mode
• Azure Databricks, only through SDK local mode

Question 5

These four are the only Targets that support ML Designer

Answer 5

Only the AML-Dedicated targets!!!

• Computer Clusters
• Serverless Compute
• Computer Instances
• ML Kubernetes

Question 6

Compute Instances:
- Require this (minimum)
- Are suitable for these kind of Models
- When they run out of disk space (120GB), do this … before you do this…

Answer 6

• Azure ML SDK or CLI v1 minimum
• Small Models less than 1GB in size
• Use the terminal to clear out 1-2 GBs of space … Stop/Restart the Compute Instance

Question 7

Both Compute Clusters and Compute Instances have these capabilities

Only Compute Clusters have these capabilities

Answer 7

Both:
* Single Node Cluster (Compute Clusters are 1 or more nodes)
* Automatic Cluster Mgmt and Job Scheduling
* Supports CPU and GPU

Compute Clusters Only:
- Multi-Node Clusters
- Autoscaling on each job submission

Question 8

Three Cost Management options for Compute Targets

Answer 8

• Offload Compute Cycle Management to Serverless Compute (preview)
• Make sure Compute Cluster minimum nodes go back to 0
• Enable Idle Shutdown for Compute Instances

Question 9

The five Unmanaged Compute Targets

Answer 9

• Remove VMs
• Azure HDInsights
• Azure Databricks
• Azure Data Lake Analytics
• Kubernetes

Surprisingly, Local Compute, Apache Spark Pools and Azure Batch are considered managed…