Using CI/CD to Automate Model Retraining and Redeployment

Having explored various statistical tests in Chapter 9, courtesy of diverse open source libraries on Databricks and their integration with MLflow, we will now focus on an integral component of MLOps on Databricks. In this chapter, we will look at how Databricks unifies DevOps, DataOps, and ModelOps all in a single platform.

In this chapter, we will cover the following topics:

• Introduction to MLOps
• Fundamentals of MLOps and deployment patterns

Let’s understand what MLOps is.

Introduction to MLOps

MLOps serves as a multidisciplinary approach that merges the principles of DevOps, ModelOps, and DataOps to facilitate the end-to-end life cycle of ML projects. It aims to streamline the transition from model development to deployment, while also ensuring effective monitoring and management. In this framework, we have the following:

• DevOps: This focuses on the continuous integration and deployment of code, aiming for quicker releases and more reliable software
• ModelOps: This specializes in managing ML models, ensuring they are effectively trained, validated, and deployed
• DataOps: This deals with data management practices, encompassing everything from data collection and preprocessing to storage and analytics

MLOps improves the performance, stability, and long-term efficiency of ML systems. There are two primary risks that MLOps can help mitigate for your use case and industry:

• Technical risks: These result from poorly managed models...

Fundamentals of MLOps and deployment patterns

To effectively manage MLOps, it’s essential to first familiarize ourselves with its underlying terminology and structure. This includes understanding the roles and responsibilities associated with various operational environments – namely, development (dev), staging, and production (prod). Let’s dissect what these environments signify in a practical MLOps framework.

Within any ML project, there are three pivotal assets:

• Code base: This serves as the project’s blueprint. It contains all the source code related to data preprocessing, model training, evaluation, and deployment.
• Data: This includes the datasets that are used for training, validating, and testing the model. The quality and availability of this data directly influence the model’s efficacy.
• Trained model: This is the culmination of your ML workflow, a model that has been trained, evaluated, and prepared for inference.

...

Understanding ML deployment patterns

The ultimate goal of any ML project is to get our ML model into production. Depending on what kind of use case we are catering to and how sophisticated our ML engineering team is, there are two broad ML deployment approaches:

• The deploy models approach
• The deploy code approach

Let’s understand these approaches one by one.

The deploy models approach

The model deployment workflow adheres to a structured methodology, beginning in a development environment where code for training the ML model is both crafted and refined. After the model undergoes training and the optimal version is ascertained, it is formally registered within a specialized model registry. This is followed by a battery of integration tests to evaluate its performance and reliability. Upon successfully passing these assessments, the model is first elevated to a staging environment for further validation. Once it meets all requisite criteria, it is then...

Summary

In this chapter, we covered the basics of MLOps, the different deployment approaches on Databricks, and their reference architectures.

Selecting a model deployment approach should be based on your team’s proficiency in implementing DevOps processes for ML projects. It’s important to acknowledge that there is no universal solution as each approach we have discussed has its own advantages and disadvantages. However, it is possible to create a customized hybrid ModelOps architecture within the Databricks environment.

By considering your team’s strengths and expertise, you can determine the most suitable deployment approach for your project. It’s essential to assess scalability, maintainability, ease of deployment, and integration with existing infrastructure. Evaluating these aspects will help you make an informed decision and optimize the model deployment process.

In Databricks, you have the flexibility to tailor your ModelOps architecture...

Further reading

Please go through the following sources and their links to learn more about the topics that were covered in the chapter:

1. The Big Book of MLOps: bit.ly/big-book-of-mlops
2. MLOps Stack on GitHub: https://github.com/databricks/mlops-stack
3. Damji, J. S., Wenig, B., Das, T., and Lee, D. (2020). Learning Spark (2nd ed.)