Is this the right deployment strategy?

By now, you have become familiar with the several types of prediction service deployments: batch, stream, real-time, and edge. You know also that there are different strategies to follow depending on your system's context. For this exercise you will imagine a colleague at work comes to ask you for help. He is building an automated MLOps system that will:

Run predictions in batch
Run the prediction every day at midnight
Be a part of a non-critical system
The predictions will not be delivered back to any user nor system
The performance of the predictions will be evaluated

Your colleague suggests that a blue/green deployment strategy is the way to go. What can you suggest to your colleague?

This exercise is part of the course

Fully Automated MLOps

View Course

Hands-on interactive exercise

Turn theory into action with one of our interactive exercises

Start Exercise

This exercise is part of the course

Fully Automated MLOps

IntermediateSkill Level

4.9+

74 reviews

Start Course for Free

In this first chapter, we motivate the use of MLOps in an industrial setting. You’ll learn about its importance in supporting the generation of value in businesses. You’ll also recap the ML stages, focusing on how MLOps enhances these. At the end of the chapter, you’ll explore a reference architecture for a fully automated MLOps system. You will then use this architecture to explore components important for any MLOps system and a starting point for the rest of the course.

Exercise 1: Introduction to fully automated MLOps Exercise 2: Hidden technical debt must be paid Exercise 3: Hidden tech debt in ML is different Exercise 4: MLOps lifecycle stages Exercise 5: Fully automated MLOps Exercise 6: To automate or not to automate...Exercise 7: Building for scale Exercise 8: Reference architecture: Fully automated MLOps Exercise 9: Reference architectures in IT and MLOps Exercise 10: MLOps architecture components

In this chapter, you will continue your exploration of the critical components that make up a fully automated MLOps system. First, you’ll examine the importance of orchestration in MLOps and how it helps to ensure the efficiency and scalability of ML pipelines. After this, you’ll examine the different deployment strategies in MLOps and learn how to choose the right strategy for your system. Finally, you’ll learn about CI/CD/CT/CM and how it complements orchestration and best practices to achieve full automation in MLOps systems. With these lessons under your belt, you will be better equipped to build a fully automated MLOps system that is efficient, accurate, and reliable.

Exercise 1: Automated experiment tracking Exercise 2: What should we keep track of?Exercise 3: The great experiment tracking debate: Which side are you on?Exercise 4: What are we tracking here?Exercise 5: The model registry Exercise 6: Do not throw the model over the fence!Exercise 7: The ML model lifecycle Exercise 8: This is a job for the model registry!Exercise 9: The feature store in an automated MLOps architecture Exercise 10: An advantage of using a feature store Exercise 11: The roles of the feature store Exercise 12: The metadata store Exercise 13: Metadata in MLOps Exercise 14: Model retraining with the metadata store Exercise 15: The role of the metadata store in monitoring MLOps

In this chapter, you’ll dive into the exciting world of automation and learn how to design more resilient and efficient ML systems. You'll start by understanding the fundamentals of automation in MLOps systems and then move on to discover the power of design patterns and fail-safe mechanisms. You'll also learn how to implement automated testing in MLOps systems and how to use hyperparameter tuning to optimize your models and workflows. By the end of this chapter, you'll be equipped with the skills and knowledge necessary to build and manage fully automated MLOps systems that are both efficient and reliable.

Exercise 1: Introducing full automation and best practices to MLOps Exercise 2: Automation in MLOps workflows Exercise 3: Degrees of automation in MLOps Exercise 4: The automation, monitoring, incident response pattern Exercise 5: Is this a design pattern?Exercise 6: Retrain, rollback or input?Exercise 7: Automated testing in MLOps Exercise 8: Garbage in - Garbage out... Apply data tests!Exercise 9: Testing: Traditional vs. MLOps Exercise 10: Testing: Data vs. model vs. pipeline Exercise 11: Automated hyperparameter tuning Exercise 12: Parameter or hyperparameter?Exercise 13: The role of hyperparameter tuning in MLOps Exercise 14: Automating hyperparameter tuning

In this final chapter, you’ll delve into the crucial components of an automated MLOps architecture. From understanding automated experiment tracking and the model registry to exploring the feature store and the role of the metadata store, this chapter is designed to equip you with a comprehensive understanding of the intricacies of a fully automated MLOps system. Whether you're a seasoned ML practitioner or just starting out, this chapter will provide you with the knowledge and skills necessary to design automated MLOps workflows.

Exercise 1: Orchestration in MLOps Exercise 2: Why should we modularize?Exercise 3: Orchestration: dev vs. prod Exercise 4: DAGs in MLOps workflows Exercise 5: Orchestration aspects in MLOps Exercise 6: Automation in MLOps deployment strategies Exercise 7: What kind of deployment is this?Exercise 8: Is this the right deployment strategy?

Current Exercise

Exercise 9: Keeping a zero downtime with a blue/green deployment Exercise 10: CI/CD/CT/CM in Fully Automated MLOps Exercise 11: How does CI/CD/CT/CM contribute to MLOps?Exercise 12: Continuous monitoring & continuous training Exercise 13: CI/CD/CT/CM in MLOps Exercise 14: Summary