1. MLOps explained end-to-end
Welcome back again! In this video, we will look at the MLOps life cycle, which we will then discuss in more detail in the following videos.
2. The broad MLOps life-cycle
Machine learning operations or MLOps consists of three major stages: design, model development, and operations, sometimes called productization.
Importantly, MLOps is not a one-way street but rather a roundabout where we use the insights we made in production to continuously refine our models to keep our application well-aligned with the business goals.
3. The MLOps life cycle - design
The design phase is the most critical stage of the MLOps life cycle because the success of the later application depends very much on a well-designed plan on how to move forward.
This stage includes gathering business objectives and requirements, prioritizing machine learning applications, inspecting the availability of the required data, and, importantly, designing the overall architecture.
4. Designing MLOps applications
In the design phase, we need to collect the relevant business metrics and define the technical or statistical metrics to assess our models. Importantly, we need to align these metrics. Otherwise, the application might not fit the organizational needs or customer requirements well and might not contribute to its performance. For an example of a mismatch, imagine, a machine learning application for a pension fund. If the aim is to generate steady returns but minimize risks, it is not a good idea to use the usual statistical metrics that are symmetric in the sense that they treat gains and losses similarly. It is an important management task to ensure that all critical metrics are aligned.
5. The MLOps life cycle - development
The next phase is called development, or model development. Here we will develop our machine learning models, evaluate, and test them. For this purpose, we need to bring the data identified in the design phase in the proper form. This is also an essential part of the development phase, which must happen before model development.
6. Developing MLOps applications
We usually think of the development phase when talking about machine learning. But training and tuning machine learning models is only one part of our MLOps life cycle. To successfully maintain and adapt our MLOps application over the long term, we must implement traditional software engineering best practices and dutifully document and track all that we have done here.
7. The MLOps life cycle - operations
As the name says, the third and final operations phase involves deploying and operating the model. Once in production, we must observe whether the models perform as expected.
8. Deployment
Bringing our models into production, or as we say deploying them, includes some critical steps between the development. These include assuring that we can always go back and restore previous versions of our application. We also need to ensure that the model will run and produce the same results on the deployment server as on the training machine.
9. Operating MLOps applications
The application is running and generates business value; now, we need to monitor and, importantly, regularly re-train the models. Why is re-training so critical?
The quality of productive models will usually deteriorate over time. The reason for that is simply that our world is changing. That means that the real-world input data will increasingly deviate from the original training data. It also means that the relationship between the input data, our predictors or features, and the outcome of interest will look different. If we suddenly witness high inflation in developed countries, that will likely lead to other credit decisions than in the previous low-inflation world.
10. The MLOps life cycle
This chapter will shed much more light on the MLOps life cycle
11. The MLOps life cycle - details
by looking more in detail at ten important steps of the three main phases.
12. Let's practice!
Now let's practice what we learned!