MLOps, The What, Why and How.

As companies are becoming more and more data-driven with more focus on building machine learning teams and creating models, it’s important to understand the challenges that ML projects present in order to properly address them with good standardized practices. The process of building a machine learning solution is complex and involves many steps that you may be familiar with: Collecting and processing raw data, Analysing the data, Processing the data for training, constructing, train, and test the model, detect bias and analyze performance to validate the model, and finally deploy and monitor the model. It’s clear how work-extensive the entire process can get when you need to repeat it multiple times as business needs and data are constantly changing. This is where MLOps comes in.

MLOps, The What and Why.

Fundamentally, MLOps is the standardization and streamlining of machine learning life cycle management, it adopts DevOps principles and applies them to ML projects to help uniting development cycles followed by data scientists and machine learning engineers with that of operational teams to help ensure continuous delivery of high-performance machine learning models. MLOps is crucial in a sense It makes it significantly easier to deploy and maintain your machine learning solutions by automating most of the hard parts so teams can keep up with the latest in machine learning technology and deploy new models quickly while monitoring their performance and ensuring that they behave as expected.

MLOps is a general underlying process that informs and automates all the steps of an ML project life cycle, helping organizations to reduce risk and introduce transparency as the data pipeline becomes visible to anyone in the organization who wants to drill down and understand how deployed ML model work. Moreover, MLOps is an important component to scaling ML capabilities, it allows teams to go from using one model to thousands of models while prioritizing continuous delivery and best model quality, which could positively impact the business and add value to it.

MLOps Setup

There mainly three established levels of MLOps depending on the maturity of ML process representing the various stages of automation and velocity of training new models given new data or implementation. The different setups are, from the one involving no automation we call it manual implementation, to the fully automated one automating ML and CI/CD pipeline.

MLOps level 0 : Manual implementation, No MLOps

At this basic level of maturity, everything is manually implemented. All ML process steps from data preparation to model deployment are manually performed. In other words, any change in the data leads the data science team to update the model and repeat the whole ML process which is usually driven by experimental code produced in notebooks and converted to scripts. The process has infrequent releases and doesn’t follow any CI/CD practices which make the ML process totally disconnected from the operations. In practice, level 0 has two main stages: The experimental stage, which involves the machine learning side of the workflow, and the Deployment stage, which deals with integrating the model into the business and maintaining it. The following diagram shows the workflow of this process.

svg Manual ML steps to serve the model as a prediction service - Google Cloud

MLOps level 1 : Continuous Model training and delivery

This level contains pipelines for automatic training of the deployed model as well as for speeding up the experimental process. we introduce automated data and model validation steps to the pipeline, as well as pipeline triggers and metadata management to automate the process of using new data to retrain models in production. Experiments results are tracked and both training code and resulting models are version controlled. The whole system is divided into reusable, composable components with modularized code. The following diagram shows the workflow of this process:

svg ML pipeline automation for continuous model training and delivery - Google Cloud

The feature store concept is introduced to replace the data storage in the previous level, it standardizes the data to a common definition that all processes can use. Both experiment and deployment pipelines will use the same input data as it held to the same definition.

In order to retrain models on new data, a trigger initiates the automated training pipeline based on different conditions. the process can be manually initiated, initiated on a specific schedule, based on a change in the data patterns or when performance drops below a certain benchmark.

MLOps level 2 : CI/CD for pipelines

At this level, the system is thoroughly able to test pipeline components before they are packaged and ready to deploy. This will ensure continuous integration of pipeline code along with continuous delivery of pipelines. Furthermore, organizations can now keep up with significant trends in the data and be able to create new models, new pipelines, and the latest machine learning trends and architectures all thanks to fast pipeline creation and deployment combined with CI/CD.

svg CI/CD for automated ML pipeline - Google Cloud

As Machine learning technology has known a significant progress in the last decade, it seems that the infrastructure is catching up too.

References :

  1. MLOps: Continuous delivery and automation pipelines in machine learning - Google
  2. What Is MLOps? - O’Reilly Media
Written on February 28, 2021