Skip to main content

Why you need to start thinking about MLOps

(Image credit: Image source: Shutterstock/PHOTOCREO Michal Bednarek)

Businesses increasingly solve complex problems with data science. Access to very large data sets, accelerated advances in ML research fields, and inexpensive computing power are driving an AI-fueled transformation across industries. In a crowded market where consumers can have anything at any time, ML/AI applications that prevent fraud, mitigate churn, serve product suggestions in real-time, and manage predictive maintenance on infrastructure can be the critical differentiator. Yet as AI/ML projects come into the mainstream, businesses are finding just how hard it is to go from data science to business value.

It’s dangerous for any company to think of these AI-driven wins as coming for free. ML projects can quickly explode technical debt without a dedicated operations practice. And worse, anywhere from 75 percent to 90 percent of ML projects don’t even make it to production, according to various estimates. Here’s another instance of the Pareto principle being a constant in life, after death and taxes: it’s everything before and after the development of the model that slows the ML process down and prevents it from seeing the light of day.

MLOps fills this gap. As the name suggests, MLOps brings the rigor and efficiency of DevOps into AI/ML practices. Its goal is to create continuous development and delivery (CI/CD) of ML intensive applications.

ML systems have a special capacity for incurring technical debt because they have all of the maintenance problems of traditional code, amplified by an additional set of data and ML specific issues. To address this, MLOps optimizes the entire lifecycle of ML, from data collection and preparation, model development, testing, and deployment, to monitoring, governance, and business metrics. The goal of MLOps is to create an environment where ML technologies can generate business value by rapidly, frequently and reliably building, testing, scaling, and releasing ML technology into production.

Getting to production

Companies like Amazon and Uber achieved dominance by leveraging AI/ML as a core competency. In the vast majority of companies where AI/ML is not part of the DNA, building an AI culture is a journey, but practical steps exist to make it feasible. Companies can invest in MLOps to ensure ML projects see the light of day and quickly generate value. Here are some ways MLOps enable efficiency and shorter time to market:

MLOps helps you develop models smarter


Typically, ML teams spend the bulk of their time on everything besides the data science. MLOps lets the scientists get back to science by automating the time consuming tasks of data collection, preparation, training, testing, and deployment.


Producing accurate Machine learning models requires massive amounts of computational power and storage. Data scientists need to be able to train a model and deploy to production without spending valuable time on scaling up infrastructure. When it comes time to train a model on larger data sets, and later when deploying to production on real world data, the time suck becomes acute. MLOps improves your team’s workflow with approaches like serverless technologies, allowing you to write code that automatically translates to auto-scaling production workloads. Serverless computing enables your team to use computing power on-demand as the need arises and manages resources efficiently. Your data scientists and your business team will thank you.

MLOps helps you deploy models smarter

Once a model is deployed to production on real world data, everything changes. Models need to be tuned or retrained, data changes unexpectedly, streaming data requires preparation and scaled up computing power. Like DevOps, MLOps seeks to speed up time to value by rolling out models seamlessly and continuously, by building models production-ready from the beginning, and automating many of the tasks involved in packaging and production.

MLOps helps you manage models smarter


It’s obvious why control and compliance is a showstopper for an industry like fintech. But recent scandals like this illustrate just how essential governance is for any model. MLOps provides access control, traceability, and audit trails to minimize risk and ensure regulatory compliance.


Deploying a model to production is only the beginning of its lifecycle. A key capability of MLOps is its ability to monitor a model for concept drift. The sudden and radical changes in global patterns resulting from Covid-19 is an extreme example of how important it is to be able to quickly react to rapid change with fresh real-time data. MLOps practices help teams collaborate, iterate, and reuse models without effort duplication, while auto-scaling resources.

MLOps: The new competitive frontier

A competitor that delivers customer value faster eventually outpaces you. The software development world has long codefied the ingredients necessary for delivering business value with predictable velocity. The product building community expresses it as a system of moving from hypothesis to validation and iteration as quickly as possible. Getting to business value with ML applications is no different in principle, though the practice is more complex.

In the best market conditions, companies look for ways to generate value quickly and cut costs. The Covid-19 crisis has placed many companies in an existential fight for survival, and getting AI applications to market will be a critical differentiator when the dust settles on this challenging period. Business teams can ill afford inaccurate predictions, application downtime, or delays in bringing models and to production. From this lens, decreasing time to market with MLOps is not a nice-to-have, it’s mission critical.

The upheavals of 2020 should also give companies pause when they consider the operational challenge of monitoring and maintaining a single ML model, let alone several. The Covid-19 pandemic period caused ML models to go awry with rapidly shifting conditions without relevant datasets to inform them. Proper MLOps practices speed up intervention when models degrade.

Achieving impact with MLOps

Prepping data, building models, getting them out of the testing environment and into the real world is a complex challenge that involves coordination across many roles in an organization.

As AI/ML applications become pervasive across industries, IT Leaders will find that embracing MLOps is a critical investment that drives value for the entire organization.

Yaron Haviv, CTO Iguazio