When enterprises begin to pursue data science and machine learning (ML) at scale, it becomes the responsibility of chief information officers and data science leaders to limit areas of liability. Managing ML models is more than just open source tools and brittle workflows though, companies also need to adopt a mindset that’s focused on continual evaluation and optimization.
This calls for increasing collaboration and productivity and shortening model delivery schedules from quarters to days or weeks. Companies also need a higher level of reliability, governance and interpretability of the data, features and algorithms at scale. I’d also be remiss if I didn’t acknowledge that we’re in the middle of a global pandemic that is increasing the pressure for businesses to tighten margins and replace head count with artificial intelligence (AI) that can maintain productivity even when employees get sick.
Introducing machine learning operations (MLOps) within your company can help to stabilize and scale ML processes. MLOps address your team’s ability to keep everything up and running, adopt an experimentation mindset and tackle the challenge of deploying more than one or two models at once.
However, there is also a major cost of adopting a MLOps practice into your business: time, money and change management that centers the right goals to succeed. This includes tactical steps like labeling data and upskilling existing employees to combine expertise in data science, software engineering and IT operations.
There are two options where technology can help companies adopt MLOps practices more easily: building or buying an ML platform.
Option 1: Build In-House
Running ML at scale is a well-known challenge for tech — already Apple, Uber, Airbnb, Netflix, and several others have committed countless hours and resources to create and maintain proprietary ML platforms. Take a look at Apple’s Overton, Netflix’s ongoing ML platform research and development, Airbnb’s Bighead, and Uber’s Michelangelo — just to name a few. These companies understand the cost of custom development, but have come to appreciate the benefits of an ML platform.
However, building a solution takes years and headcount. Airbnb, for instance, took around three months figuring out what to build — and about four years to build it. By the end, they used numerous open source technologies and still had to fix the gaps in the path to production by defining their own services and user interface. This required expertise in supporting multiple frameworks, feature management, model and data transformation, multi-tenant training environments and much more. Similarly, Uber has been working for five years on their platform and Netflix started more than four years ago.
And unless you are a big tech firm, you’re likely struggling with hiring AI talent. When I built a data science team from scratch, I needed to decide: hire classically trained data scientists or hire domain-specific experts and upskill. I chose to upskill. I’m not alone. In a 2020 PwC survey on AI, when asked how they plan to handle the shift to more AI, 46 percent of organizations said they’re rolling out AI upskilling and 38 percent are implementing credentialing programs.
There are certainly benefits to building in-house, like supporting lesser-known use cases of AI that aren’t yet supported. And, for technology companies whose core capabilities include building, maintaining and open sourcing specific tools, it absolutely makes sense to build in-house. But this is a core competency you’ll want to intentionally choose to build a team around.
Option 2: Buy a Commercial ML Platform
When buying a platform, you’re saving more than the initial cost to build. It also saves integration costs for custom brittle workflows and provides dedicated external support. You’re also lowering the ongoing cost of onboarding new employees to proprietary software — reducing time to productivity. But weigh the benefits of buying an enterprise-grade platform against the costs, which includes adopting new workflows instead of building for those you already have and removing some of the prolific tools your teams may prefer to use.
Often emerging platforms offer fewer stable workflows initially, so you’ll need to evaluate the scope of integrations and supported workflows before making a purchase decision. Consider your MLOps needs in terms of collaboration, reliability, governance, auditing and interpretability of the data, features, and algorithms. Not all platforms will support the entirety of your ML operations or your company’s unique needs. Evaluate carefully.
Ultimately, emerging technologies promise to level the playing field without needing to hire, build and maintain additional teams to support the change. Enterprise-grade ML platforms allow your teams to focus on what differentiates your business. This allows you to achieve the efficiencies and collaboration of the big tech companies, while gaining an edge with your unique domain expertise.
Most companies recognize, or are beginning to recognize, that ML is the future. Companies are committing more headcount and resources to the challenge. But is it enough? There are major costs of adopting a MLOps practice into your business. New commercial ML platforms, however, give companies a way to leap forward in their drive to deliver ML in production and at scale. The right ML platform will support the implementation of your MLOps strategy to meet the challenge.