Machine Learning Operations – MLOps

Early concept made possible with technological advancement

The origin of artificial intelligence (AI) as a computer science field that aims to simulate every aspect of human intelligence by using computing machines dates to 1956, when John McCarthy et. al organised the Dartmouth Summer Research Project on Artificial Intelligence ¹ . That workshop was the seed of a unified identity for the field and a dedicated research community that led to numerous breakthroughs during the subsequent decades ² . Despite the rapid growth of different aspects of artificial intelligence, the field could still boast no significant practical success and nearly every computer program that was built to interact with humans was codified with a defined set of rules, only allowing rudimentary displays of intelligence in specific contexts and with capabilities limited to specific tasks. Ruled-based systems, also known as expert systems, cannot be designed to address complex problems where the programming of the set of rules that defines the body of knowledge is impractical or impossible.

It was not until the nineties that the AI focus started drifting from expert systems toward a new paradigm built upon the idea that programs should be able to learn, so that AI could start displaying its potential ³ . It is in this new paradigm where modern AI establishes its roots, and the field of machine learning (ML) is born, being the “field of study that gives computers the ability to learn without being explicitly programmed,” as stated by Arthur Samuel. Machine learning algorithms learn through data exposition, usually via an iterative process or an ensemble of statistical samplings of the data. A key feature of ML is that its predictions improve with experience and with the use of more relevant data, at least up to a certain point. Thus, by learning through practice, instead of following defined sets of rules, machine learning systems deliver better solutions than expert systems in numerous cases, and more complex problems start being accessible.

Catalyst for digital business

70% of the globe’s GDP will have gone through some form of digitisation by 2022, and by 2023, investments in Direct Digital Transformation will amount to $6.8 trillion ⁵ . As companies are undergoing their journey of digitalisation today, the use of ML is a key feature in automating, predicting, planning and personalising their product. However, the integration of ML within the business chain comes with new challenges that have a tremendous impact on business. Some of the questions that companies are nowadays facing are not related to how to build ML models, but rather to which built models are in use, what they are doing and whether the used data reflects the state of the world. Although the answer to these questions might seem simple when compared with the complexity of the ML algorithms, they are usually overlooked, bearing negative effects on business ⁶ . To align models with business needs and to generate business value, it is therefore essential not only to build the ML model but also to deal with dataset management, monitoring and deploying models and building processes that are shareable and repeatable throughout an organisation. To provide a solution to those issues, a set of best practices have been codified into a new field, Machine Learning Operations, or MLOps.

Therefore, MLOps is a set of practices for the operationalisation of ML models that aim to build, deploy and monitor ML applications and that facilitates the collaboration and communication between data scientists and operations professionals quickly and reliably. Some of the MLOps capabilities are ⁷ :

• MLOps allows the unification of the release cycle for machine learning and software application releases.
• MLOps enables automated testing of machine learning artefacts, e.g., data validation, ML model testing and ML model integration testing.
• MLOps enables the application of agile principles to machine learning projects.
• MLOps enables supporting machine learning models and datasets to build these models as first-class citizens within CI/CD systems.
• MLOps reduces technical debt across machine learning models.
• MLOps must be a language-, framework-, platform-, and infrastructure-agnostic practice.

Seven core principles

According to Microsoft’s Machine learning DevOps guide ⁸ , seven core principles should be considered when adopting MLOps for any ML-based projects: 

1. version control code, data, and experimentation output – to ensure reproducibility of experiments and inference results 

2. use multiple environments – to segregate development and testing from production work, as shown below.

3. manage infrastructure and configuration with infrastructure-as-code – for consistency between environments 

4. track and manage machine learning experiments – for quantitative analysis of experimentation success and to enable agility 

5. test code, validate data integrity, model quality – to test the experimentation code base 

6. machine learning continuous integration and delivery – to ensure that only qualitative models land in production 

7. monitor service, models, and data – to serve machine learning models in an operationalised environment 

It is worth mentioning that the interpretation of these Microsoft principles should be flexible, i.e., they are not a set of rules that must be adopted when designing an ML project. Specifically, the second core principle, use multiple environments, can be omitted most of the time during development, even though it is relevant to follow it for functional testing of applications and APIs. As an illustration, in our solution, we adopt only a reduced set of those principles.

In this work, a proposed MLOps solution is presented in terms of infrastructure configuration, data preprocessing workflow and end-to-end model development workflow. The aim is to showcase how MLOps principles can be brought into practice using cloud computing resources from Azure. The proposed MLOps solution is based on the above-mentioned core principles, where the focus in this paper is on principles 1,3,4 and 6.

Proposed MLOps solution 

The core components of the proposed MLOps solution can be summarised in terms of infrastructure configuration, data preprocessing workflow and end-to-end model development workflow, which will be presented below.

Infrastructure configuration

• Infrastructure-as-code: terraform 
• Code repository: Azure DevOps repo 
• Data repository: Azure data lake with a Delta Lake storage layer 
• Model repository: MLflow 
• Model hosting server: Azure Pipeline Docker 

Verdict

This work proposes an MLOps solution for scalable and repeatable end-to-end ML implementation. The solution is set up using Infrastructure-as-code and contains semi-automated steps for data preprocessing and end-to-end machine learning development workflows. As noticed throughout the development, MLOps increases the quality, simplifies the management process and automates the deployment of ML models in a large-scale production environment. Thus, it becomes easier to align models with business needs.

MLOps creates a wide array of benefits, namely:

• MLOps orchestrates the entire development process
• MLOps monitors data drifting
• MLOps leverages agile methods
• MLOps promotes truly reusable components
• MLOps versions both data and models
• MLOps alleviates comparison between models and artifacts

In contrast, there are also some limitations worth mentioning, such as a lack of a common definition of data within the data pipeline, and the dependence on certain tech stacks that limit the generalisation of certain procedures.

All in all, MLOps is a set of principles for establishing a common working framework when implementing ML solutions to meet business operationalisation needs. However, the practical implementation of MLOps depends specifically on each case. Experiment with it and pick the solution that fits your business needs the best.