ModelOps is the latest evolution in the Machine Learning and AI industry designed to get the data science field over its biggest hurdle.
The percentage of data science models that makes it into production is around 20%, so why are so few models making it into production and how does ModelOps help?
ModelOps and DevOps
ModelOps gets its origin from DevOps, with DevOps being a ground breaking approach to quickly develop, deploy and update software and custom applications. This is achieved in DevOps by bringing together into one team everything that is required to develop software quickly. It also involves the IT, security, user experience experts, compliance and more. One element of DevOps that is also important is the shared success metrics, the entire team works together towards a shared set of objectives.
Now ModelOps or Model Operations, is a version of DevOps but for the development and deployment of data science models. For most organisations, they have a team or third party team of data scientists that build models but require internal IT resources to deploy and implement these models into the organisation. This process can take months and involve a lot of work.
ModelOps is the systems and processes involved in developing and deploying data science models. It involves bringing together data scientists, data engineers, application owners, infrastructure teams, testers and technology.
This new ModelOps team needs to set clear objectives, identify what models need to be deployed, when and work together. The principles of clear communication and knowledge sharing is key as this is one of the main drivers that reduces time to deployment.
The benefits of ModelOps
The ModelOps approach is about empowering your data scientists and focusing every resource required to deliver models to do just that. ModelsOps is about creating a team with the goal of putting models into production, rather than just having data scientists who’s goal it is to develop models.
When implemented correctly ModelOps is about enabling your data scientists with the tools and resources they need to provide your organisation with data science models that drive business value. ModelOps, should lead to quicker deployment of models and more models making it into production.
The challenge ModelOps faces is how to scale it. The approach still requires large teams of people to deploy data science and that creates resource constraints. Teams looking to deploy models still have capacity constraints. On average it takes a team of 6 people 10 working days to deploy a model on existing technology platforms. That is 60 days per deployment. This can become very costly for organisations looking to deploy and update multiple models per month.
So what is the answer?
An alternative approach
The challenge facing ModelOps is one that faces the entire data science community. ModelOps takes a broken process and throws people at it. On average ModelOps teams are around 6 people including Data Engineers, ModelOps/DevOps, Developers, IT infrastructure and Data Scientists.
Having this team in place is a key step that many organisations can take but it comes with a high cost and assumes you can easily get these skills in a very competitive marketplace.
An alternative approach is to look at make data science and model deployment easier.
How Lumen helps ModelOps
Lumen is a data science delivery platform and allows Data Scientists to deploy models more efficiently thus saving time and money. Lumen brings the automation required to streamline the deployment process.
Lumen negates the need for a team of developers, once the platform is operational you can deploy models quickly and without a team of ModelOps engineers and support. This allows you to focus your efforts on testing and updating a wider range of existing models and looking to explore new use cases and ensure that your investment in data is producing effective returns.