The enterprise industry is buzzing over a new development and operational model that brings together the existing disciplines of DevOps and Machine Learning (ML). Dubbed machine learning operations or "MLOps," the goal is to establish an end-to-end process for the design, development and management of powerful new ML-based software products.
What is MLOps?
While still in its infancy, the movement has captured the attention of everyone from data scientists and software engineers to experts in the field of artificial intelligence and machine learning. One of its chief proponents is MLOps.org, which has identified a number of unique capabilities that MLOps brings to traditional software engineering. These include:
- Unification of the release cycle of ML and software applications.
- Enablement of automated testing of ML artifacts like data validation, model testing and integration testing.
- Application of agile principals to ML projects.
- Support of ML models and datasets as first-class assets in CI/CD systems.
- Reduction of technical debt across ML models.
- Establishment of MLOps as an agnostic practice across languages, frameworks, platforms and infrastructure.
DevOps and MLOps
The ultimate goal for MLOps, says Nvidia’s Rick Merritt, is to establish a set of best practices to allow organizations to successfully deploy and operate a wide range of ML and other AI-empowered applications. By basing MLOps on existing DevOps models, backers are hopeful that these new products and services can transition smoothly into established digital business models. (Read also: DevOps Managers Explain What They Do.)
In essence, it provides an easy on-ramp into the established data environment for data scientists, analysists, engineers and others who specialize in the data curation, automation and related functions utilized by ML-driven programs.
Why Do We Need MLOps?
Ideally, this will establish an AI-ready infrastructure that can quickly onboard key elements of the MLOps software stack, such as a repository of AI models, an automated ML pipeline to manage data sets and experiments, as well as a fleet of software containers – most like based on Kubernetes – to simplify job processing. But don’t get the idea that this is a simple matter of cutting and pasting a new model onto the old. Working with disparate data sets will require careful labeling and tracking, while strong sandbox and repository management will be needed to ensure a stable test environment. (Read also: Is AI Going to Replace Computer Programmers Anytime Soon?)
While MLOps does bring advantages to devtest, it really starts to shine once the project hits the production phase. According to algorithmia.com, MLOps has the ability to integrate organization, management and monitoring into a single programmatic process that incorporates hardware orchestration, language and SDK integration, container management, model versioning and a host of other functions, including key security processes like access control and encryption.
Improving the Business Model
Ultimately, these benefits make their way to the business model by:
- Accelerating time-to-value, perhaps from months down to minutes.
- Optimizing team productivity through integrated workflows and role specialization.
- Improving infrastructure management to better suit business outcomes.
- Protecting business assets and continuity.
From an organizational perspective, MLOps represents the transition from the “era of artisanal AI” to the “application of engineering disciplines to automate ML model development, maintenance and delivery,” according to a recent Deloitte report entitled “MLOps: Industrialized AI”. This addresses one of the key problems with current efforts to implement AI in the enterprise: it tends to exist as the domain of a few star data scientists who exercise broad creative control over the technology and its application. While this may result in a few innovative solutions, it limits the ability to scale them to enterprise levels.
At the same time, they are hampered by legacy infrastructure that cannot support rapid, consistent, streamlined development. (Read also: Machine Learning: 4 Business Adoption Roadblocks to Consider.)
Intelligence at the Speed of Business
Using automation and standardized processes, MLOps encourages experimentation and rapid delivery, in part by democratizing AI across a wider spectrum of the knowledge workforce. With better data organization tailored for machines, new techniques can be implemented quickly, even autonomously, to adjust business processes and models to changing environments. As well, feedback loops can ensure that outdated models are decommissioned while newer more productive ones move to the forefront.
Of course, this transition cannot happen until the AI skills gap at most organizations is addressed. Machine learning can propel a business model to new levels of performance, but without the necessary expertise to guide and manage it, well, it can wreak quite a bit of havoc.
Intelligent or not, technology has shown itself to be a key determining factor in the success or failure of any complex endeavor, but this should not overshadow the fact that humans remain the most important asset in the enterprise. No matter what form it takes, MLOps will only be as good as the people who run it.