Azure ML Operationalization

Streamlining and expediting a data scientist's CI/CD workflow leveraging prebaked functionalities of agentless Azure DevOps Pipelines, Azure Functions, and Azure Service Bus Queues.

Key aspects:

• Azure DevOps Pipelines
• Azure Service Bus
• Azure Functions
• Azure Machine Learning

Core dependencies:

• Python 3.6.8
• Azure Functions Core Tools
• Azure CLI
• Azure ML SDK + Contrib

Main benefits:

• Avoiding relying on build agents that might be preoccupied to run notebooks
• Improved telemetry in the DevOps UI to debug notebook failures
• Support for testing and validating against custom images
• Quicker run-times from avoiding slow-downs associated with build agents

Looking forward:

• Azure ML Pipelines are an alternative to the Service Bus + Azure Function combination. There is more of a story around shipping this module with an ML Workspace. There have been concerns over how third parties would have to initialize the Azure Function Application and Service Bus Queue with the current implementation.
• Event Grid or a similar event-driven trigger on completion of a Run would allow for a more integrated, platform-level replacement to the current code-injection method.
• Making this a more agnostic solution will require thought on our current, in-house release.json solution and how a more generally applicable solution might be achieved. Other aspects like Git Integration can be streamlined by requiring all repositories to be imported into Azure DevOps to allow DevOps to handle authentication.
• The benefits of agentlessness are felt most when all phases in a pipeline are agentless. Therefore, phases like Get Changed Notebooks should also be moved to an agentless architecture whenever possible.

The File Directory

• RunNotebookFunctionApp

• run_notebook_service_bus

• __init__.py
• function.json

• handlers

• azureml_handler.py
• devops_handler.py
• file_handler.py
• notebook_handler.py

• deployment_pipeline.yml
• automlcli_example_pipeline.yml
• requirements.txt

deployment_pipeline.yml

This agent-based pipeline prepares the deployment environment, unit-tests, bundles, and finally deploys the Azure Function Application:

- stage: Deploymentjobs:
- job: pool:
vmImage: 'ubuntu-16.04'steps:
- task: UsePythonVersion@0
- bash: pip3.6 install -r requirements.txt
- bash: pytest
- task: ArchiveFiles@2
- task: AzureFunctionApp@1

automlcli_example_pipeline.yml

This agentless pipeline runs a given notebook with a given imagine on a specified compute target, and provides telemetry results as test runs in the DevOps UI.

- stage: Validationjobs:
- job: pool: serversteps:
- task: PublishToAzureServiceBus@1

requirements.txt

This file controls the dependencies required for the Azure Function. This is mainly the Azure ML SDK, its associated dependencies, and support for HTTP requests, file system manipulation, etc.

...
azureml==0.2.7
azureml-contrib-notebook==1.0.43
azureml-core==1.0.43
azureml-pipeline==1.0.43
azureml-sdk==1.0.43
...

When deploying the function app, these are injected into the package, as seem in the deployment YAML file:

pip3.6 install -r requirements.txt

run_notebook_service_bus/function.json

This file specifies the location of main function (e.g. __init__.py), as well as the Service Bus binding for the function.

run_notebook_service_bus/__init__.py

This script holds all the pythonic logic of the application. The main function is short, favoring a helper function to handle the distinct job types:

start_build_pipeline()

Fetches the repository of interest, creates a new Experiment SDK Object, and submits a set of notebook Runs to that object after injecting try-catch statements to facilitate callbacks to the DevOps pipeline.

update_build_pipeline()

Updates the DevOps Test Runs based on results from Azure ML Compute, and checks to close the pipeline in all Runs are completed.

handlers/azureml_handler.py

This script handles all Azure ML SDK-related logic, managing Azure ML Workspace-related tasks:

fetch_exp()

This function authenticates with the ML Workspace with a Service Principal connection, fetches the Workspace, and then fetches and returns a new Experiment.

fetch_run_config()

This function generates a RunConfiguration based on the pipeline parameters, specifying such things as the ComputeTarget and CondaDependencies. More flexible of Run configurations can easily be implemented.

submit_run()

This function submits a new Run with configurations based on the pipeline parameters.

fetch_run()

This function fetches a Run by its RunID tag specified by the DevOps Test Run.

fetch_exp_status()

This function determines the status of the pipeline by fetching all Runs and checking their status

handlers/devops_handler.py

This script handles all DevOps related tasks:

post_pipeline_callback()

This function, along with its helper functions, closes the DevOps Pipeline via the DevOps API.

post_new_run()

This function, along with its helper functions, creates a new DevOps Test Run via the DevOps API.

patch_run_update()

This function, along with its helper functions, updates a DevOps Test Run via the DevOps API.

post_run_attachment()

This function, along with its helper functions, uploads a DevOps Test Run Attachment via the DevOps API.

post_run_results()

This function, along with its helper functions, adds a DevOps Test Run Result with result telemetry via the DevOps API.

get_repository()

This function, along with its helper functions, downloads a zip of a specified DevOps repository and a specific branch and subfolder.

handlers/file_handler.py

This script handles all file IO related tasks:

prepare_staging()

This function unzips a repository archive and places it in the staging directory for build_snapshot() to access.

add_pip_dependency()

This function injects pip dependencies into the Conda file specified in the pipeline parameters.

add_notebook_callback()

This function adds try-catches around each code-block of a notebook with callbacks to re-trigger the Azure Function when the notebook is finished running.

Note: this is a not an "ideal" solution from an architectural perspective, but a more platform-level, agnostic approach (e.g. with Event Grid integration for triggering the Function) is currently out of scope.

remove_notebook_callback()

This function removes try-catches around each code-block of a notebook after the notebook has been executed in Azure ML Compute so that results can be displayed cleanly in Azure DevOps.

fetch_requirements()

This function finds a notebook's release.json file to determine dependencies, requirements, and other run parameters.

build_snapshot()

This function moves repository files from the staging directory into the snapshot directory, keeping the snapshot directory as small and light-weight as possible to run a specific notebook.

handlers/notebook_handler.py

This class, Notebook, handles all code manipulation for notebooks to be fed to Azure ML Compute:

inject_code()

This function adds a collection of lines of code at the front or back of a collection of specified code cells.

inject_cell()

This function adds new code cell for pre- or post-execution scripts.

scrub_code()

This function removes all lines of code injected by inject_code from a collection of specified code cells. It also removes injected code cells from the beginning and end of the notebook.

Glossary

DevOps Pipeline Variables

Name	Description	Example Value	Where To Find
do.organization	DevOps Organization's Name	example-organization-name	DevOps Organization's Mnemonic Name
do.project	DevOps Project's Name	example-project-name	DevOps Project's Mnemonic Name
ex.compute	Azure ML Compute Target's Name	example-compute	Machine Learning Workspace's Assets
ex.image	Azure ML Compute's Target Image	mcr.microsoft.com/azureml/base:intelmpi2018.3-ubuntu16.04	Azure ML SDK's RunConfig Module
fx.azureSubscription	Azure Subscription for Function App	ExampleSubscription(a1234567-89bc-0123-def4-abc56789def)	Function App's Overview
fx.name	Function App's Name	example-function-app	Function App's Mnemonic Name
gh.repo	Location of GitHub Repository	https://github.com/example/repo	GitHub Repository's Overview
rp.condaFile	Repository's Conda File Location	src/example-notebooks/environment.yml	Repository's File Directory
rp.notebooks	Repository's Notebooks to Run	one.ipynb,two.ipynb,three.ipynb	Repository's File Directory
rp.version	Repository's Commit/Branch of Interest	"bb7ad65dbc727ec09fe0613d51ce8585087de1b1", "master", "dev", etc.	GitHub Repository's Overview
sb.connection	Service Bus Queue's Connection String	Endpoint=sb://example.servicebus.windows.net/...	Service Bus Queue's Shared Access Policies
sb.name	Service Bus Queue's Name	example-queue-name	Service Bus Queue's Mnemonic Name
sp.client	Service Principal's Application (client) ID	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)	App Registration's Overview
sp.password	Service Principal's Password	32 character alphanumeric string (e.g. A/fb0...)	App Registration's Client Secret
sp.tenant	Service Principal's Directory (tenant) ID	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)	App Registration's Overview
ws.name	Machine Learning Workspace's Name	example-ws-name	Workspace's Mnemonic Name
ws.resourceGroup	Workspace's Resource Group's Name	example-resource-group-name	Workspace's Overview
ws.subscription	Machine Learning Service Workspace's Subscription ID	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)	Workspace's Overview

DevOps System Variables

Name	Description	Example Value
system.AccessToken	DevOps Bearer Token	Long Opaque String
system.HostType	DevOps Pipeline Type	"build"
system.JobId	DevOps Pipeline Identifier	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)
system.PlanId	DevOps Pipeline Identifier	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)
system.TaskInstanceId	DevOps Pipeline Identifier	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)
system.TeamProjectId	DevOps Pipeline Identifier	GUID (e.g. a1234567-89bc-0123-def4-abc56789def)