MLOps on Azure

Data Scientists and ML practitioners usually are good at building models and performing complex statistical analysis but may not be good at software engineering skills required to put them into production. Also, a simple DevOps approach is not sufficient to address an ML project. It needs to take into account models, data, experiments, runs and artifacts produced. By way of this module we explore how model deployment is being done now vs how could the ideal state be.

Azure Machine Learning (AML) which is a cloud service offering has built-in MLOps. It can easily take care of the different phases of the Machine Learning Development Cycle (MLDC) and also offers complete integration right out of the box with popular ML frameworks like Scikit-learn, Pytorch, Tensorflow and MXNet. This module will provide a deeper look into how AML provides support for all the different phases of the MLDC.

Using Azure Pipelines, ML Engineers can automatically build and test ML projects. It seamlessly integrates CI (Continuous Integration) and CD (Continuous Delivery) and ships the ML model and supporting code to a target of choice. Azure Pipelines also facilitates CT (Continuous Testing) by helping automate the build, test and deploy processes efficiently.

Docker is a containerization mechanism to package the code and the underlying runtime as a Docker image to enable modularity and portability. The aim is to remove the occurrences of ‘it worked on my laptop’ by containerizing all the required elements of the model. Kubernetes is a container orchestration platform which helps in scheduling and orchestrating all the workloads among the different containers.

Understand how Azure ML helps in deploying the ML models by 1. ensuring that the source code is controlled, 2. Creating reproducible training cycles 3. Data and Environment management 4. Creating and maintaining ML pipelines 5. Utilizing model registry to track models and experiments 6. Evaluate and Validate the model 7. Monitor and retrain it as needed

MLflow is an open-source platform which can run on Azure (or any other public cloud or on-prem infrastructure) which also helps in managing the ML lifecycle. Understand the 4 major components

1. MLflow Tracking
2. MLflow projects
3. MLflow Models
4. Model registry and see how MLflow can be used in conjunction with Databricks on Azure to create an efficient MLOps framework.

Kubeflow is another open-source framework for MLOps which is gaining popularity very quickly. It is a framework which works on top of the Azure Kubernetes Service (AKS) and helps in orchestrating the different phases on MLDC without a lot of overhead of managing low level Kubernetes APIs. This module teaches how KFserving, Knative and other sub components of Kubeflow work on Azure to create another awesome MLOps framework.