As the adoption of artificial intelligence (AI) and machine learning (ML) continues to proliferate, applying MLOps best practices has become increasingly crucial to ensuring scalability, efficiency, and reliability. Recently, more regulations for real-world AI applications have been introduced and discussed—such as the European AI strategy—making model governance a new top priority for companies of all sizes.
Model governance is a supplementary component to MLOps that supports AI compliance and traceability by improving visibility and control over ML deployments. Model governance inside MLOps processes—i.e. MLOps governance—enables better performance for production ML systems both from an engineering perspective, as risk is minimized and quality maximized, and from a legal perspective, as regulatory compliance can be guaranteed and easily demonstrated.
This article defines model governance and MLOps governance, explains why MLOps governance is important, and explores how to implement a governance framework with respect to defining both regulatory requirements and MLOps processes.
AI/ML model governance is the set of processes that ensures traceability over model deployments. The degree of ML governance and MLOps integration varies depending on the number of models in production and the degree of regulation in the business domain. When more models are in production, ML model governance is fully embedded inside MLOps practices. One way to understand this is by conceptualizing MLOps governance as being embedded in ML systems (see Figure 1.)
Figure 1: The intersection of ML governance and MLOps based on the number of models and regulations (source: ml-ops.org)
MLOps governance guarantees the full integration of ML governance processes inside the end-to-end model lifecycle. Governance processes become seamless to reuse and simpler to extend for multiple ML production pipelines. During development and experimentation, MLOps governance ensures reproducibility and validation by providing artifact tracking and resource sharing. During deployment and in production, it ensures observability, security, and cataloging by providing the right documentation and system access for each stakeholder.
MLOps governance introduces automated processes for tracking, monitoring, validation, documentation, and for versioning all ML artifacts, including complete data, models, code, and pipelines. Commonly, the governance processes are categorized as either data management or model management. Model management encompasses the last three artifacts (models, code, and pipelines,) since data is likely already to be subject to pre-existing organizational paradigms.
MLOps model governance introduces fine-grained levels of control and visibility into how ML models and pipelines operate in production, with different access and insights given to different stakeholders. This managed traceability offers a number of benefits:
The ability to perform the above reliably can make a real difference to the success of any ML initiative, especially those that are long running. The impact extends beyond guaranteed compliance and better engineering practices to offer benefits with regard to reputation management and better model performance for end users.
A model governance framework refers to the entirety of the systems and processes that exist to support the model governance requirements for all production ML models. In the early stages of ML adoption, model governance is often performed manually and may lack unified tools and processes. While this may be necessary initially, as the team and its processes grow more mature, manual processes do not provide a robust enough framework to ensure proper governance or allow ML scalability to multiple models.
Defining an ML governance model is not an easy task, since the discipline is relatively new and regulatory requirements change frequently.
There are two phases to establishing a model governance framework:
The guidelines provided for each phase, offered below, can be used as a sanity check for anyone implementing or reviewing their model governance model.
Different ML use cases are typically expected to require different regulatory compliance. Creating or reviewing existing regulatory processes for model governance is thus a necessary step for any new production ML model pipeline.
This process involves understanding the ML use case category and identifying the individuals responsible for overseeing governance processes. Additionally, determining the necessary governance policies for the use case, including PPI (personal identifiable information,) sector-specific regulations, and regional regulations is necessary. Integrating these policies within the MLOps platform and engaging and educating stakeholders are also vital steps in establishing regulatory compliance. Finally, monitoring and refining processes ensure that the model remains in compliance with regulations over time.
MLOps governance can be divided into data governance and model governance. Data governance is typically integrated inside broader IT policies. Ideally, model governance would also be integrated into such policies. However, in reality MLOps extends standard DevOps, meaning that model governance requires its own processes and policies—which are specific to machine learning—to be set up.
A model governance framework can be implemented by introducing these MLOps advanced functionalities:
Model monitoring ensures efficient tracking of all assets and model performance, with smart alerts available to notify the appropriate stakeholders at the right time of any issues that arise.
An end-to-end MLOps platform such as Iguazio ensures easy implementation of the latest best practices in developing, deploying, and managing machine learning models and pipelines, including supporting the seamless embedding of model governance processes.