Path8.Mod1.a - Intro to DevOps Principles for ML

Question 1

Three Processes we combine when discussing MLOps

Answer 1

• ML Workloads: Exploring data analysis, Feature Engineering and Model Training/Tuning
• Dev: Software development; planning (model reqs and metrics), create (model training and scoring script), verify (check code and model quality) and package (staging the solution)
• Ops: Release (deploy to prod), Configure (IaC - Infrastructure as Code) and Monitor (Metrics tracking and performance)

Question 2

Describe CI/CD w.r.t. MLOps

Answer 2

• Continuous Integration: Create and verify activites; Refactoring notebooks to Python code, Linting, unit testing
• Continuous Delivery: Steps to deploy to staging then prod, automating as much as possible.

Question 3

It is preferrable that Training Data is not stored in Source Control. Store it in a database or data lake and retrieve it directly from the source via a Data Asset (T/F)

Answer 3

FALSE. Retrieval should be through a Datastore since the Datastore will store all credentials for connecting to the storage source.

Otherwise TRUE you should NOT store your training data in Source Control.

Question 4

BRP IRA

DevOps tools in Azure DevOps and GitHub and what we get out of them

Answer 4

Azure DevOps:
* Azure Boards - sprint planning and tracking boards
* Azure Repos - code repositories
* Azure Pipelines - CI/CD

GitHub:
* Issues - sprint planning and trackgin boards
* Repos
* Actions - automated workflows

Question 5

c ad/gh s d p t n

Best Practices for MLOps Repository Structure (what to put in which folders)

Answer 5

.cloud: cloud-specific code (ARM templates, etc)
.ad/.github: Azure DevOps or GitHub artifacts like YAML pipelines to automate workflows.
src: contains any code (Python) used for ML workloads
docs: Markdown files or other documentation describing the project.
pipelines: Azure Machine Learning pipelines definitions.
tests: contains unit and integration tests
notebooks: contains notebooks, mostly used for experimentation.

Question 6

What to use:
- for building Azure ML Pipeline templates
- for Local Development

Answer 6

Python SDK or YAML + CLI, and VS Code respectively

Question 7

non func mod unit wkfl

Steps for refactoring Notebooks to production Python code

Answer 7

• Removing all nonessential code
• Refactoring code into functions
• Combining related functions in Python scripts to modules (.py files)
• Creating unit tests for each Python script
• Create a pipeline to group scripts into workflows that can be automated

Question 8

Three ways to create an Azure ML Pipeline (aka Pipeline Job)

What YAML would go into pipeline-job.yml

Answer 8

• Define steps in a Python script (Job-based classes)
• With Components in Designer
• YAML using CLI v2 commands

For the pipeline-job.yml approach you’d define:
- type: pipeline
- jobs for transform-job and train-job