DevOps for AI will help you to understand how to build the Continuous Integration and Continuous Delivery pipeline for a ML/AI project. 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.
This template contains code and pipeline definition for a machine learning project demonstrating how to automate the end to end ML/AI project. The build pipelines include DevOps tasks for data sanity test, unit test, model training on different compute targets, model version management, model evaluation/model selection, model deployment as realtime web service, staged deployment to QA/prod and integration testing.
This reference architecture shows how to implement continuous integration (CI), continuous delivery (CD), and retraining pipeline for an AI application using Azure DevOps and Azure Machine Learning. The solution is built on the scikit-learn diabetes dataset but can be easily adapted for any AI scenario and other popular build systems such as Jenkins and Travis.
- One-time run: These tasks run only for the first time that the build pipeline runs. They will programatically create an [Azure ML Service Workspace](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#workspace), provision [Azure ML Compute](https://docs.microsoft.com/en-us/azure/machine-learning/service/how-to-set-up-training-targets#amlcompute) (used for model training compute), and publish an [Azure ML Pipeline](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-ml-pipelines). This published Azure ML pipeline is the model training/retraining pipeline.
3. The Azure ML Retraining pipeline is triggered once the Azure DevOps build pipeline completes. All the tasks in this pipeline runs on Azure ML Compute created earlier. Following are the tasks in this pipeline:
- **Train Model** task executes model training script on Azure ML Compute. It outputs a [model](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#model) file which is stored in the [run history](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#run).
- **Evaluate Model** task evaluates the performance of newly trained model with the model in production. If the new model performs better than the production model, the following steps are executed. If not, they will be skipped.
- **Register Model** task takes the improved model and registers it with the [Azure ML Model registry](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#model-registry). This allows us to version control it.
- **Package Model** task packages the new model along with the scoring file and its python dependencies into a [docker image](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#image) and pushes it to [Azure Container Registry](https://docs.microsoft.com/en-us/azure/container-registry/container-registry-intro). This image is used to deploy the model as [web service](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#web-service).
4. Once a new model scoring image is pushed to Azure Container Registry, the Azure DevOps Release/Deployment pipeline is triggered. This pipeline deploys the model scoring image into Staging/QA and PROD environments.
- In the Staging/QA environment, one task creates an [Azure Container Instance](https://docs.microsoft.com/en-us/azure/container-instances/container-instances-overview) and deploys the scoring image as a [web service](https://docs.microsoft.com/en-us/azure/machine-learning/service/concept-azure-machine-learning-architecture#web-service) on it.
5. The deployment in production is a [gated release](https://docs.microsoft.com/en-us/azure/devops/pipelines/release/approvals/gates?view=azure-devops). This means that once the model web service deployment in the Staging/QA environment is successful, a notification is sent to approvers to manually review and approve the release. Once the release is approved, the model scoring web service is deployed to [Azure Kubernetes Service(AKS)](https://docs.microsoft.com/en-us/azure/aks/intro-kubernetes) and the deployment is tested.
