Watch on YouTubeThe modern ML stack does not exist yet because teams have different requirements and the field lacks shared conventions.
MLOps can mean the workflows a team uses to develop and productionize models, or the infrastructure that runs ML systems, and those are different concerns.
Joe Reis argues that durable data and ML education should focus on processes, life cycles, and fundamentals instead of tools that may change.
Mike Del Balso and Joe Reis discuss why the modern ML stack has not settled into a clear set of composable services. Data teams have more established practices, while ML teams still describe similar requirements with different terms. Mike uses feature stores as an example: one team may want transformations in custom Java services, another may want them in model code, and another may simply ask for a feature store. That variation makes consolidation difficult. They also separate MLOps workflows from the infrastructure that runs ML systems. Joe connects this uncertainty to his book, The Fundamentals of Data Engineering, arguing that durable guidance should explain processes and life cycles rather than chase changing tools. The conversation ends with the value of in-person MLOps Community meetups, where people exchange practical use cases, learn from one another, and find hiring opportunities.
Mike Del Balso says the modern ML stack is still being built. The field has many vendors and products described as data platforms for ML, along with several kinds of stores, which makes the boundaries hard to explain. He expects a clearer stack to emerge as vendors work together and make their systems simpler to describe. One sign of maturity would be seeing most practitioners draw broadly similar diagrams when asked to describe the stack.
Joe Reis says the data domain is more mature because it has had time to consolidate its practices and processes. Earlier systems such as Informatica described much of the work needed for data movement, while ELT changed the arrangement without making the overall process impossible to describe. He sees ML as more open-ended and less complete. Mike adds that ML users often have genuinely different requirements, which makes it hard for a small set of products to cover the field.
Mike gives feature stores as a concrete example. An e-commerce, security, or gaming company may want feature transformations in a custom Java application, inside model code, or in a feature store. These teams may describe their needs differently even when one solution could meet the underlying requirements. Mike expects teams to recognize that their use cases are similar over time, allowing fewer solutions to cover more situations.
Mike says education is part of the problem. He describes a company that built three variations of a feature store and called two of them feature stores. People at that company had inherited different ways of thinking from those earlier systems. Newer companies can sometimes move faster because they have no historical artifacts and can adopt a clearer description of their ML stack from the beginning. Joe frames this as a lack of convention rather than a lack of available technology.
Mike says people use MLOps to mean different things. In one usage, it means the workflows a team follows while developing and productionizing machine learning. In another, it means the infrastructure that runs ML systems. Some products provide infrastructure and try to support a workflow, while others combine infrastructure components mainly to help a team move faster. Treating these as the same thing makes the field harder to describe.
Joe explains that he wrote The Fundamentals of Data Engineering to work out what data engineering is and consolidate the knowledge around it. When asked how a book can cover a fast-changing field, he says the answer is to focus on what will not change. He avoids centering the book on tools and instead writes about processes, life cycles, and fundamentals. Writing also exposes gaps in his own understanding, while reviewers help identify blind spots.
The speakers describe their first in-person MLOps Community meetup as a useful source of accumulated learning after a long period without events. Conversations covered real products and edge cases, including computer vision systems for Amazon Go-style retail checkout. Mike says these meetings expose him to a wide range of ML use cases. Demetrios Brinkmann wants future local gatherings to stay lightweight, with people meeting for a happy hour rather than needing to organize presentations.
Mike says meetups offer more than informal conversation because attendees can find jobs and hiring leads. He mentions that Tecton is looking for people to build feature stores, while other companies are hiring for different roles. Demetrios adds that participants share information about open positions and connect one another with people working on relevant problems. The value comes from open conversations about what people are building and where they are from.
"Sometimes they mean I'm in charge of the workflows that my team will have as they develop and productionize machine learning, and sometimes MLOps to them just means the infrastructure that runs your ML systems."Mike Del Balso08:41