Where does MLOps fit in the Digital Transformation journey?
Disclaimer — As a PM tasked with building ML products, I may express bias towards AI’s impact; if you don’t believe me, let this challenge your perspective
Recently, I’ve gotten the chance to zoom way out of the MLOps problem space to the 30,000 ft view of enterprise digital transformation, the multi-decade initiative to close the efficiency gap between digitally-native and legacy enterprises. My goal for doing so is to revisit the reasons why we need MLOps in the first place and where does it fit in the larger digital transformation roadmap.
What are the 4 main themes I see organizations focusing on in their digital transformation agendas?
- New ways of working
- Understand your business and customers through data analytics
- Rethink your business processes
- Operationalize AI
Let me summarize the following themes followed by explaining MLOps role to accelerate the AI value produced by data teams operationalizing AI.
New ways of working
Legacy businesses scaled teams and delivered value in a centralized, linear way. Development projects happened in siloes and teams were burdened by excess coordination costs and dependency management issues. This waterfall approach added more risk (not less) to projects and slowed innovation to crawl. Overtime a new iterative way of delivering value to the business and its customers won out.
These new ways of working helped reshape enterprise topologies and emphasized value-creation at the edge — where the users are.
Understand your business & customers
If one thinks of organizations as a complex systems and their customers as a relationships, then the incontestable need to study the inter-workings of the organization require the capturing of data. Collecting and curating data serves as the foundation for all forms of analytics to take place.
TLDR: Data is the major key 🔑. Data teams are working to break down the analytical siloes and empower everyone to understand the business. Making data and insights accessible leads to better business decisions across the enterprise at every level.
Rethink your business processes
As organization size grows so does complexity. Organizations aren’t static either, they continue to change leading to inefficient spaghetti-like processes where no one wins. Manual processes which can be automated and human judgement codified go through robotic process automation (RPA) or custom app development. Most transformation programs try to Marie-Kondo their operations by simplification or removal.
Applying technology and analytics to optimize business processes is here to stay.
Operationalize AI
With new ways of working, an understanding of the past and present state of the business, and simplified business processes one can start applying ML to optimize operations or create new offerings once thought unattainable. The reality for many small medium businesses is that ML is difficult. Framing business problems as ML use cases, curating quality datasets, figuring out the tooling, deploying, monitoring, retraining, the list goes on! All of the ML failures in 2010–2020 have led to low trust and confidence in operationalizing AI.
Enter MLOps — a synthesis of DevOps practices with ML-specific processes to manage models end to end. As mentioned in ‘Rethink your business processes’ section, DevOps practices enable the continuous release of high quality software through automation and observability. Sound DevOps teams built trust in their organizations. MLOps teams hope to bring the same practices to manage ML artifacts (data, code, and models) with reliability, reproducibility, and scale.
TLDR: MLOPs major key 🔑 — ensure frictionless ML services and development using automation to all
How will it evolve?
If you combine the 4 main digital transformation themes mentioned above and look at a timeline, you’ll start to see the 2010’s were all about data-driven companies. Data collection and storage became inexpensive. SAAS data platforms helped catalogue, transform, and manage enterprise data. Data could be interrogated through ad-hoc reports and dashboards. Data teams focused on optimizing the pipeline, improving the latency, and scaling analytics across the organization. The 2010’s saw data become accessible.
In the 2020’s I see the following trends taking on a more important role:
- MLOps becomes a standardized trusted set of practices for the enterprise amongst all modalities and edge cases
- Engineers focus more on data quality, data labelling, and model management problems
- Executives ask the right questions and build cultures of experimentation
- ML-as-a-service emerges empowering anyone to improve and scale predictions solving business problems
As companies continue to digitize their operating models, MLOps will play a major role in making ML accessible to all.
