MLOps #01: An Introduction to MLOps levels & its life-cycle
A step by step guide to MLOps levels & evolution
In these series about MLOps, I will be covering in each blog post a step of the MLOps pipeline and deep dive into it. Today, as it is the first blog, I will start with introducing MLOps and all the basics that you need to know to 🚀
Motivation
This graph illustrates that 10% of a Machine Learning project is the training code, but what creates technical debt in production ML systems is the fact that all the other parts of the ML workflow are not well taken care off, maybe for the lack of tools and knowledge about them.
MLOps comes as solution to technical debt in ML systems.
What is MLOps?
MLOps stands for Machine Learning Operations.
- It is inspired by traditional DevOps that empowers the build-deploy-monitor lifecycle for software applications.
- The only difference to DevOps is that MLOps is an ML engineering practice/process that aims to unify ML system developments (Dev) and ML system operations (Ops).
MLOps level 0
MLOps level 0 is characterised by its highly manual steps.
This graph illustrates a typical Machine Learning pipeline.
We can split the pipeline to two phases:
- Machine Learning phase that deals with data extractions & analysis, data preparation, model training, model evaluation & validation until model registry.
- Operations phase that consists of the serving of the ML model and using it in production.
MLOps level 0 characteristics
- Every step in the workflow is manual. Each step is executed manually, any the transition between steps is also manual.
- The machine learning and operations component of the machine learning system are disconnected.
- Deployment of this workflow is all about getting the model into a prediction service, typically a micro service with a REST API.
- Suitable for unfrequent ML models releases.
MLOps level 1
MLOps level 1 aims to automate the Continuous Training of Machine Learning models.
MLOps level 1 consists of automating the process of model re-training with new data from production.
Adding the human in the loop who is responsible to review the predictions that will be used in the Continuous Training.
MLOps level 1 characteristics
- Quick and flexible experimentation thanks to the automation of CT.
- Human in the loop = domain expert => Data quality insurance.
- Having Supervised metrics that are more insightful than the unsupervised metrics
- Model validation automation thanks to the supervised metrics
MLOps level 2
“MLOps level 2 aims to enable rapid and reliable updates of ML models in production using CT/CD”
MLOps level 2 is an evolution of MLOps level 1 with slight differences.
MLOps level 2 characteristics
- Piepline orchestration.
- Continuous Deployment = the automated deployment of the re-trained models resulting from CT into production.
- More complex of implementation but gaining the benefits of flexibility.
Final thoughts
To address the challenges of MLOps level 0, Continuous Integration, Continuous Training & Continuous Deployment component were leveraged that led to produce respectively level 1 & 2 of MLOps.
The choice of the MLOps level is dependent on the requirements and specification of the ML project. Therefore all the levels are valid.
Thanks for reading and see you in the next blog 👋
