End to end MLOps

We will be using the Azure DevOps Project for build and release/deployment pipelines along with Azure ML services for model retraining pipeline, model management and operationalization.

We now have 3 pipelines that we need to setup.

• Build Pipeline (azure-pipelines.yml): Runs tests and sets up infrastructure
• Retraining trigger pipeline(/template/retraining-template.json): This pipeline triggers Azure ML Pipeline (training/retraining) which trains a new model and publishes model image, if new model performs better
• Release pipeline(/template/release-template.json): This pipeline deploys and tests model image as web service in QA and Prod environment

Note

Start by cloning this repo https://github.com/trallard/MLOps_complementary And make sure to have followed the instructions on Additional setup

1. Set up Build Pipeline

1. Select your devops organization and project by clicking dev.azure.com

2. Once you are in the right devops project, click Pipelines on the left hand menu and select Builds

3. Click New pipeline to create new pipeline

   

4. On the Connect option page, select GitHub
5. On the Select option page, select the GitHub repository where you forked the code.
6. Authorize Azure Pipelines to access your git account
7. Since the repository contains azure-pipelines.yml at the root level, Azure DevOps recognizes it and auto imports it. Click Run and this will start the build pipeline.
8. Your build run would look similar to the following image
   

Great, you now have the build pipeline setup, you can either manually trigger it or it gets automatically triggered everytime there is a change in the master branch.

Note: The build pipeline will perform basic test on the code and provision infrastructure on azure. This can take around 10 mins to complete.

2. Set up Retraining trigger release pipeline

Note: For setting up release pipelines, first download the release-pipelines to your local filesystem so you can import it.

To enable the option to Import release pipeline, we must have

atleast one release pipeline so let’s create one with an empty job.

On the next screen, click on Save and then click Ok to save

the empty release pipeline.

Steps

1. Select the Release tab from the menu on the left, then click the

   New dropdown on top and click on Import Release pipeline

1. On the next screen, navigate to release-pipelines folder and

   select retrainingtrigger.json pipeline file, click import. You should now see the following screen. Under Stages click on the Retrain stage, where it shows the red error sign.

1. We would now link the variable group we created earlier to this release pipeline. To do so click on the Variables tab, then click on Variable groups and then select Link variable group and select the variable group that we created in previous step and click Link followed by Save button.

1. We want the retraining pipeline to be triggered every time build

   pipeline is complete. To create this dependency, we will link the artifact from build pipeline as a trigger for retraining trigger release pipeline. To do so, click on the pipeline tab and then select Add an artifact option under Artifacts.

1. This will open up a pop up window, on this screen:
   • for source type, select Build
   • for project, select your project in Azure DevOps that you created in previous steps.
   • For Source select the source build pipeline. If you have forked the git repo, the build pipeline may named yourgitusername.DevOpsForAI
   • In the Source alias, replace the auto-populated value with ``DevOpsForAI``
   • Field Devault version will get auto populated Latest, you can leave them as it is.
   • Click on Add, and then Save the pipeline

1. Artifact is now added for retraining trigger pipeline, hit the save button on top right and then click ok.

2. To trigger this pipeline every time build pipeline executes, click

   on the lighting sign to enable the Continous Deployment Trigger, click Save.

1. If you want to run this pipeline on a schedule, you can set one by

   clicking on Schedule set in Artifacts section.

2. For the first time, we will manually trigger this pipeline.

• Click Releases option on the left hand side and navigate to the release pipeline you just created.
• Click Create Release
• On the next screen click on Create button, this creates a manual release for you.

Note: This release pipeline will call the published AML pipeline. The AML pipeline will train the model and package it into image. It will take around 10 mins to complete. The next steps need this pipeline to complete successfully.

3. Set up release (Deployment) pipeline

Note: For setting up release pipelines, first download the release-pipelines to your local filesystem so you can import it.

Also Note: Before creating this pipeline, make sure that the build pipeline, retraining trigger release pipeline and AML retraining pipeline have been executed, as they will be creating resources during their run like docker images that we will deploy as part of this pipeline. So it is important for them to have successful runs before the setup here.

Let’s set up the release deployment pipeline now.

1. As done in previous step, Select the Release tab from the menu on

   the left, then click the New dropdown on top and click on Import Release pipeline

2. On the next screen, navigate to release-pipelines folder and

   select releasedeployment.json pipeline file, click import. You should now see the following screen. Under Stages click on the QA environment’s **view stage task”, where it shows the red error sign.

Follow the same steps for Prod Environment and select **Hosted

Ubuntu 1604** for agent pool and save the pipeline.

1. We would now link the variable group we created earlier to this

   release pipeline. To do so click on the Variables tab, then click on Variable groups and then select Link variable group and select the variable group that we created in previous step and click Link followed by Save button.

2. We now need to add artefact that will trigger this pipeline. We will add two artifacts:

   • Build pipeline output as artifact since that contains our configuration and code files that we require in this pipeline.
   • ACR artifact to trigger this pipeline everytime there is a new image that gets published to Azure container registry (ACR) as part of retraining pipeline.

Here are the steps to add build output as artifact

• Click on pipeline tab to go back to pipeline view and click Add an artifact. This will open a pop up window
  • for source type, select Build
  • for project, select your project in Azure DevOps that you created in previous steps.
  • For Source select the source build pipeline. If you have forked the git repo, the build pipeline may named yourgitusername.DevOpsForAI
  • In the Source alias, replace the auto-populated value with ``DevOpsForAI``
  • Field Devault version will get auto populated Latest, you can leave them as it is.
  • Click on Add, and then Save the pipeline

Here are the steps to add ACR as an artifact

- Click on pipeline tab to go back to pipeline view and click **Add an artifact**. This will open a pop up window
- For Source type, click on **more artifact types** dropdown and select **Azure Container Registry**
- For **service connection**, select an existing service connection to Azure, if you don't see anything in the dropdown, click on **Manage** and [create new **Azure Resource Manager**](https://docs.microsoft.com/en-us/azure/devops/pipelines/library/service-endpoints?view=azure-devops#create-a-service-connection) service connection for your subscription.
**Note:** You must have sufficient privileges to create a service connection, if not contact your subscription adminstrator.
- For Resource Group, select **DevOps_AzureML_Demo**, this is the default resource group name that we are using and if the previous pipelines executed properly you will see this resource group in the drop down.
- Under Azure container registry dropdown, select the container registry, there should be only one container registry entry.
- For repository, select **diabetes-model-score** repository.
- For Default version, keep it to **latest**
- For Source alias, keep the default generated name.
- Click Add
- Click on lighting sign to enable the **Continous Deployment Trigger**, click Save.

.. image:: https://github.com/microsoft/MLOpsPython/raw/master/docs/images/release-deploymentcitrigger.png

1. We now have QA environment continously deployed each time there is a new image available in container registry. You can select pre-deployment conditions for prod environment, normally you don’t want it to be auto deployed, so select manual only trigger here.

Congratulations, you now have three pipelines set up end to end.

• Build pipeline: triggered on code change to master branch on GitHub.

• Release Trigger pipeline: triggered on build pipeline execution and produces a new model image if better than previous one.

• Release Deployment pipeline: QA environment is auto triggered when there is a new image.

  Prod is manual only and user decides when to release to this

  environment.