ModelOps: Handsome Data Scientists Need Not Apply (Part 1)
In 2020, People magazine crowned film star Michael B. Jordan as the sexist man alive. Each year, we see a new handsome celebrity heartthrob grace the covers of People, yet some individuals within your enterprise may remain sexy year after year: data scientists. Harvard Business Review (HBR) may have put it best with their decree of data scientists being “The Sexist Job of the 21st Century”.
HBR was right. We’ve seen an explosion in the growth of Data Science roles across multiple industries. Unlike passing fads, the bottom line justifies this growth, as the first cohort of data scientists often pay for and justify the next.
Yet, individual sexienss isn’t everything.
With the lowest hanging fruit already capitalized on, scaling and optimizing the operational processes that data scientists work within and depend on becomes imperative to maintain growth of business-changing outcomes. Over time, it’s the data science ecosystem that matters most, as opposed to individual contributors.
The first data scientist and the first model within an organization can add immense value without significant overhead. When dozens or hundreds of models require new launches, maintenance and monitoring, having the right processes and tools to reduce model-time-to-market and model risk become indispensable.
Future growth and effectiveness of data science requires teams to subscribe to a set of principles and tools that help them optimize their efforts much in the same way we’ve already seen agile developer teams do in the form of “DevOps”. We call this effectiveness-boosting framework framework, “ModelOps”.
Good Looks, or ModelOps Processes & Tools?
Imagine, for a moment, that there was a direct correlation between handsome good looks, technical competency and individual performance. The more ravishingly gorgeous the data scientist, the higher performing they would be.
If good looks were a measure of how good of a data scientist you would want on your team, what would you prefer?
(1) Five Micahel B. Jordan Data Scientists
(2) Five Anonymous “Average Looking” Data Scientists that have been trained on ModelsOps best practices, tools and frameworks
If we ask enterprises that have recently started their ModelOps journey, they would unequivocally choose the latter. They have the benefit of experiencing first hand the kind of productivity gains that can be achieved with adopting only a few ModelOps best practices.
The Top 3 ModelOps principles are:
- Principle #1: Feature Store
- Principle #2: Offer models-as-a-service (MaaS)
- Principle #3: Intelligent monitoring
Let’s explore the first principle, a highly adopted feature store.
Principle 1: Feature Store
The highest value data scientists can provide to their organization is the development of new and powerful predictive models. Despite wide acceptance of the strategic importance of these activities, data scientists often disproportionately focus on prerequisite activities. An all too common situation for many organizations is that data scientists spend considerable energy on data preparation, data cleansing and data lineage analysis.
Adoption of a Feature Store enables Data Scientists to focus on their highest value activities with curated and validated features ready to be immediately used for model development. With ongoing Feature Store stewardship, trust and adoption by Data Scientists has been observed to yield productivity gains of 4X and consistently avoid costly work duplication.
“Adoption of a Feature Store was a game changer for us. Immediately, we could see that our team members were able to deliver many new projects at a fraction of the time it was taking in the past.”
~VP Data Science & Analytics @ A National Bank with $40B+ Revenue
In addition to improving productivity and reducing model time to market, the adoption of a mature Feature Store is a key component of any data governance strategy. Key capabilities to include as requirements for open source or commercial Feature Store software solutions should include:
- Searchable catalogue of features, their metadata and statistics
- Role based access to datasets
- Data validation quality assurance metrics & reports
- Version control and data providence tools
What’s Next?
Look for part two and three of this ModelOps series, which will cover the next two principles:
- Principle #2: Offer models-as-a-service (MaaS)
- Principle #3: Intelligent monitoring
If you’re unsure about your readiness for ModelOps or your current level of maturity, check out T1A’s ModelOps 7-minute survey here: to see where your team compares to your industry peers.
