Sponsored by Anaconda
Data Science Solution Center

Forrester: 7 Key Requirements For Successful MLOps Deployment


The lack of defined, repeatable process for ML model operations may be part of the reason so many do not reach production.

Businesses are looking for ways to integrate machine learning into their organization, but many find it hard to find key use cases or struggle to deploy.

According to Forrester Research, a way to improve the chances of deployment and production success is with MLOps, a repeatable process that covers the entire ML model’s lifecycle.  

SEE ALSO: How Machine Learning Is Shaping a New Manufacturing Era

In a recent study, Forrester found that 73 percent of respondents said MLOps adoption would keep them competitive, while 24 percent said it would make them an industry leader.

“Enterprises do want to leverage ML, but most don’t have all the necessary competencies. To achieve ML at scale, enterprises must achieve MLOps capabilities,” said Forrester in a study on MLOps.

Similar to DevOps, MLOps is utilized as a way of tracking the model’s development and ensuring it does not go off track or stagnate. Mike Gualtieri, principal analyst at Forrest Research, provided seven key requirements to successfully deploy MLOps in a webinar:

Provision infrastructure resources:Machine learning models require resources throughout development, but these may change as the model moves from concept, to development, to production. “Part of what an MLOps solution has to do is be able to provision those resources throughout the cycle,” said Gualtieri.

Support multiple ML model formats: Most large businesses use multiple programming languages, the same is true with ML model development. An MLOps solution needs to be agnostic to the way the model was developed, encompassing most if not all of the popular formats.

Support software dependencies as well: Models have dependencies, especially if developed in open-source. Supporting these underlying dependencies, including version control, can help a model continue to work long after initial development.

Monitor models to ensure they are not drifting or doing harm: “Models are imperfect, even on Netflix and they’re making a recommendation, it’s not always accurate,” said Gualtieri. “Models are probabilistic because they’re trained on historical data, when the environment changes it needs to be re-trained. When they are in production they need to be monitored for drift, to make sure that they’re making good decisions.”

Ability to deploy anywhere: Deployment of ML models may have to happen on-premise, in the cloud, or at the edge. An MLOps solution that does not have a singular deployment pattern will ensure production is more flexible.

Proper governance to explain and audit models:Understanding the underlying goals of a machine learning model is critical, especially if the model is used for sensitive subjects. Good data governance provides this trust with businesses, regulators, or the public.

Retrain models when new data becomes available:Once new data becomes available that may change or derail the model, teams should work to retrain it using the original data pipeline, algorithms, and code.

By following these requirements, Gualtieri said businesses will find an easier route to deployment while reducing the chance of the ML model failing to deliver on its original goals.

Most businesses interviewed for the Forrester study lack mature MLOps, with only six percent saying they have mature capabilities. The majority, 41 percent, said they have struggled to operationalize, while 21 percent were in the proof of concept stage.

That lack of defined, repeatable process for ML model operations may be part of the reason so many do not reach production.

David Curry

About David Curry

David is a technology writer with several years experience covering all aspects of IoT, from technology to networks to security.

Leave a Reply

Your email address will not be published. Required fields are marked *