Mark Treveil and the Dataiku team
MLOps – core concepts at a glance
Automate and scale machine learning processes in the company sustainably
O’Reilly, August 2021
204 pages, from € 27.99 (book, e-book)
In the Gartner Hypecycle, machine learning has meanwhile reached the plateau of productivity: In the jargon of market researchers, this means that ML-based processes meanwhile “create value” in numerous companies, that is, contribute to business success. However, this goes hand in hand with increasing demands on the management of the processes – frantic programming is no longer the order of the day when problems that cause costs arise.
The still young discipline MLOps is a new type of DevOps thought process that exclusively deals with working with models and systems of artificial intelligence. The book “MLOps – Core Concepts at a Glance”, now published by O’Reilly, delimits the term from ModelOps and AIOps. A group of authors around Mark Treveil and the Dataiku team start the work, which has been translated into German by Marcus Fraaß, to explain this new scientific discipline in detail.
As in the case of cross-platform frameworks and other comparable challenges in IT, a crucial principle also applies to MLOps: an understanding of domains is essential when dealing with the tasks at hand. The authors therefore start their considerations with the concrete positioning of an MLOps process in the IT and business process landscape of a company in order to then name the responsible stakeholders and the components of an MLOps process to be processed by them.
Since “MLOps – core concepts at a glance” is not understood as an AI or ML tutorial, all explanations are based much more on the administrative “how”. Already in the first section it is noticeable that the collective of authors consistently garnishes its explanations with anonymized examples from real ML deployments or MLOps processes.
Life cycle management
DevOps processes usually go hand in hand with managing the life cycle of the software components they support. The second part of the book therefore examines the creation and operation process of an ML system as a whole. Here, too, the content focus is again more on the legal aspects and the project implementation than on the programming.
Particularly praiseworthy is the fact that the team of authors is also dedicated to the increasingly important topic of fair and inclusive artificial intelligence. Because anyone who uses an AI system nowadays must expect that its results will also be assessed for their political correctness.
In addition, the authors address technical problems that arise in the day-to-day operation of an ML model. This applies, for example, to the permanent monitoring of model performance: an ML model must always adapt to variable input data that either result from a change in the real situation or can also be traced back to targeted attacks. In any case, these changes can cause fluctuations in the performance of the model, which can only be detected and, if necessary, corrected through monitoring and security processes.
Almost in passing, the authors also consider how the data exchange between the prototypes newly developed by data scientists and the stable ML models already used in practice can be accomplished.
MLOps in everyday life
The collaboration with the ML consulting firm Dataiku enabled Treveil to incorporate practical examples that were only anonymized with regard to company names, countries and customer data. The third and last part of the work therefore presents specific MLOps applications in various industries. In addition to analyzing and predicting the default rate of consumer credit, there are also considerations for predicting loads in the power supply system and generating recommendations for marketing systems. As in the previous chapters, the book also focuses on the tooling and the processes used to apply the ML models for these examples.