1. MLOps on Kubernetes
In our last video, let us understand how MLOPs and Kubernetes
can work together.
2. What is MLOps?
First of all, what is MLOPs? MLOps, or Machine Learning Operations, is a paradigm for deploying and maintaining ML models in productions. It is a collection of best-practice workflows with a strong focus on continuous development and constant improvement of models.
It is inspired by DevOps, which focuses on the continuous development, deployment, and operations of software solutions.
Similar to DevOps, ML models, as well as AI models or GenAI solutions, are developed and tested in isolated systems and then deployed to production.
When productive, the models are continuously monitored, their accuracy is constantly measured, and improvements through retraining may be triggered.
Data scientists, data engineers, and IT teams can work together on deployed models synchronously, hence ensuring their accuracy.
3. Implementing MLOps on Kubernetes
Now, this paradigm of MLOps maps very well to Kubernetes.
The isolated experimental systems that we need for development and testing can easily be realized by Pods and Kubernetes Storage, of course.
We can monitor our productive ML models easily by following
the lifecycle of our Pods, together with the images we have deployed.
We can also work synchronously, as a team, on model accuracy, as this is enabled from the very beginning by the architecture of Kubernetes
Several frameworks exist for MLOps on Kubernetes. The two best-known open-source solutions are mlflow and Kubeflow.
4. Kubeflow - An Overview
Let us have an overview of Kubeflow here.
Kubeflow is dedicated to making the deployment of ML workflows on Kubernetes very simple. It covers each step of the ML model lifecycle, e.g., data gathering, data wrangling, model training and testing, and deployment.
Kubeflow consists of several components which cover all of these steps. Most important, all these components work independently of each other. This enables maximum flexibility and lets us create all types of workflows easily.
We can also use Python directly to develop and deploy ML models using Kubeflow.
There is no need to use kubectl here, we can interact directly with the Kubernetes API.
5. Let's practice!
And for the last time, let us practice what we have learned in this lesson.