Image by Author

So you’ve built a machine learning model. And it achieves the desired performance on the validation dataset. You’re happy that you were able to apply your data science and machine learning skills in building this model. However, you realize that the model working fine—inside your Jupyter notebook—on your local machine is not very helpful (yet). 

For users to benefit from your model and for businesses to leverage machine learning, the models have to be deployed to production. However, deploying and maintaining machine learning models is not without challenges. A model that performs well in the development environment may fail miserably in the production environment. This could result from data and concept drifts and other factors that lead to such degradation in performance.

So you realize: for machine learning models to be useful, you have to go beyond model building. Here’s where MLOps comes into play. And today, we’ll learn about MLOps and the role of the MLOps engineer in an organization.

What is MLOps?

More often than not, you’ll come across MLOps being defined as applying DevOps principles to machine learning.

The software development life cycle (SDLC) has changed for the better with DevOps practices simplifying cross-functional collaboration between the development and operations teams. If you know someone working in DevOps, you’d have heard them talk about CI/CD pipelines, automated CI/CD pipelines, application monitoring, and more. 

Though this can be transposed to machine learning applications, there are certain challenges specific to machine learning systems. Building and operating machine learning machine learning systems is a more involved process.

So in general, you can think of MLOps as a set of best practices to build, deploy, and maintain machine learning systems.

With this idea, let's proceed to learn what an MLOps engineer in an organization does.

What Does an MLOps Engineer Do?

We said it’s possible to apply DevOps practices to machine learning systems. If that’s MLOps, then MLOps engineers are responsible for doing exactly that! 

What do we mean by that? Once the data science team has built the model, the MLOps engineer enables the successful operation of the model through the following:

• Automating machine learning model deployment
• Setting up monitoring for the ML pipeline
• Automating CI/CD pipelines to account for data, code, and model changes
• Setting up automated model retraining 
• Deciding on the level of automation required

Some Challenges Associated with MLOps 

Setting up monitoring can only help identify when things go wrong. To get more granular information on the performance of the different versions of the model, MLOps engineers often use model versioning and experiment tracking.

We mentioned that an MLOps engineer sets up model retraining with a required level of automation. Let’s try to understand the challenges associated with it. 

Once an application is deployed to production, it’s possible that the data that the model uses—in production—is drastically different from the data that it was trained on. As a result, such a model will perform poorly, and we often have to retrain. 

MLOps engineers also handle the retraining and automation of the retraining process by factoring in performance degradation, the frequency at which data changes, and the cost of model retraining. 

Wait, I do MLOps. But I’m not an MLOps engineer.

In some startups, you may have machine learning engineers who also wear the hat of the MLOps engineer. While in some others, you have DevOps and Backend engineers who wear multiple hats. 

MLOps at a big tech company can be very different from MLOps at an early stage startup. The level of MLOps automation may also vary from one organization to the other.

If you’ve been working at a startup where you hold ownership for the end-to-end machine learning pipeline—from model training to monitoring and maintaining machine learning systems—you’re already an MLOps engineer, too.

Are you excited in exploring this challenging role of an MLOps engineer? Let’s summarize the skills you’ll need.

MLOps Skill Set and Tools: An Overview

MLOps engineers usually have strong ML, DevOps, and data engineering skills.
 

Image by Author

• Machine Learning skills: Programming, working knowledge of machine learning algorithms and frameworks, and domain knowledge
• Software engineering skills: Querying and working with databases, testing ML models, Git and version control, frameworks like FastAPI
• Devops foundations: Proficiency in tools such as Docker and Kubernetes 
• Experiment tracking: Familiarity with experiment tracking frameworks such as MLflow
• Orchestrating data pipelines: Setting up and automating data pipelines using tools such as Prefect and Airflow
• Cloud infrastructure: Familiarity with cloud infrastructure providers such as AWS, GCP, and infrastructure-as-code tools (IaC) like Terraform

Learn MLOps

If you are you interested in learning more about MLOps, here’s a list of resources to help you get started:

• MLOps Zoomcamp by DataTalks.Club: The MLOps zoomcamp from DataTalks.Club is a free course to learn all things MLOps – from model building to best practices for deployment and monitoring. You’ll learn by building a project to put together all that you’ve learned.
• MLOps Specialization on Coursera: The Machine Learning Engineering for Production (MLOps) specialization from DeepLearning.AI. The specialization (with four courses) will teach you how to build production-grade machine learning systems. 
• MLOps GitHub Repositories: A curated list of repositories to level up in MLOps.

Wrapping Up

In this article, we covered the overarching responsibilities of the MLOps engineer in an organization and the key MLOps skills. 

As mentioned, not all engineers who do MLOps are called MLOps engineers. We also addressed how the level of MLOps automation and the actual day-to-day work depends can vary from one organization to the other. 

Like any other role, succeeding as an MLOps engineer requires soft skills such as effective communication, collaboration, and strategic problem solving. That said, if you’d like to try your hands at being an MLOps engineer, happy MLOps!
 
 
Bala Priya C is a technical writer who enjoys creating long-form content. Her areas of interest include math, programming, and data science. She shares her learning with the developer community by authoring tutorials, how-to guides, and more.