Introducing ModelOps to the Organisation: What It Is and Its Benefits
Industry analysts including Gartner and Forrester have long noted that many organisations are failing to capitalise on their investment in analytics. Generally speaking, this results from a focus on model development and data science, however this then results in a struggle to integrate the models into business operations — the action that actually unlocks the value from analytics.
ModelOps is a framework or practice that has emerged to address this challenge and is inspired by the success of DevOps. Its focus is operationalising analytics, i.e. taking models from development to production, and therefore transforming modelling from an academic exercise to an economic benefit. Effectively, it activates the value of analytics by applying data science to decision-making within the organisation.
I have often heard ModelOps described as ‘sophisticated model management’. However, it is much broader than model management because it is supported by a wide range of technology, from data to decisions. It is also known as MLOps, DeepOps or AIOps and simply put is a framework that helps organisations take models from development to production effectively.
To better understand the complexities involved and what ModelOps looks like in practice, I’m going to cover the aspects that should be addressed in a series of articles. These articles cover the benefits, organizational framework, the supporting technologies and the sophistication levels of ModelOps. The content described draws on experiences helping organisations of many sizes, across many industries, assessing and implementing their ModelOps frameworks. As well as conversations with peers and research from within the industry.
The benefits of ModelOps
There are four key benefits of ModelOps:
1. Faster time to value, by streamlining the analytics lifecycle and reducing the time between development and deployment;
2. Better and more justifiable business outcomes, by bringing more governance to the analytics lifecycle and continuously monitoring the performance and business impact of the models deployed;
3. The ability to scale analytics through automation and repeatability means doing more with the same resources, through more effective collaboration between the business, data scientists, operations and IT.
4. Embedding analytical insight into every business decision and customer touchpoint, driving better, real-time and automated decision-making.
Of course, it is possible to develop and deploy analytical models without ModelOps. However, its use makes the entire process much more efficient, improving the speed, volume and governance of models being deployed in production. Additionally, the question of where ModelOps fits with DevOps and DataOps often arises when speaking with organisations. Combining the three allows organisations to take full advantage of their analytics investment and genuinely operationalise analytics, even though both DataOps and DevOps have effects well beyond this.
It is worth pointing out that ModelOps is not relevant for all organisations. Organisations will only get the full benefits of this approach if they are using predictive analytics. Those doing historical reporting, data exploration and off-line statistics should therefore take time to identify if new use cases for predictive analytics are valuable. If so, they can use ModelOps when developing these use cases to maximise their efficiency and ability to become more sophisticated in the future.
Principles of ModelOps
It is also important to note that ModelOps is not simply DevOps for models. Models are different from traditional software components. Model development needs specialised analytics skills and is an experimental process. Model deployment is also extremely complex. Models need to evolve over time to reflect changes in market conditions and the underlying assumptions about the data. They therefore need to be continuously monitored and retrained to avoid any degradation in their performance and it is necessary to establish processes to continuously monitor and retrain them.
Once organisations have decided that they can benefit from ModelOps, they need to think about how to use the approach. ModelOps is NOT a one-size-fits-all approach. The level of sophistication needed is guided by the organisation’s business objectives and its existing analytics capabilities and cultural readiness. However, even for organisations that need to reach the most advanced approach, a stepped progression through the stages will improve adoption. Crucially, using the right level of ModelOps sophistication will help the organisation to gain value from its analytics and reduce the risk of failure. We will go into more detail of what these sophistication levels look like later in a future article.
Developing a holistic approach
What does this mean in practice? It means that a holistic approach, combining people, process and technology with a supportive culture, is needed to successfully implement ModelOps. Organisations should define their process based on their business strategy, enabling this with a supportive culture. This will allow the organisation to define a consistent technology stack to facilitate the process and reduce the chances of silos and duplication of work.
The next article in this series will discuss this approach in more detail.
