Serving models remained the most common learning goal among people joining the MLOps Community, despite attention shifting toward monitoring, feature stores, and data-centric AI.
2
The MLOps ecosystem grew quickly in 2021, with many companies funded, acquired, spun out, or releasing new versions, which made the field harder to follow.
3
Vishnu Rachakonda argues that teams should build their data and analytics foundations before applying more advanced machine learning.
Summary
Vishnu Rachakonda and Demetrios Brinkmann review the MLOps Community and ecosystem at the end of 2021. They discuss the community's growth, including the gap between total signups and active members, and reflect on podcast conversations with Cody Coleman and D. Sculley. Demetrios describes how those interviews changed his view of the data engineering work behind production machine learning. The conversation then covers jargon, tool awareness, regulation, acquisitions, funding rounds, spinouts, events, community projects, and the community blog. Vishnu says the growth of the ecosystem can be overwhelming, even for people working in MLOps, and recommends starting with the problem rather than choosing a tool first. Looking ahead, he wants to understand analytical and data engineering foundations better. Demetrios is most interested in meeting more community members through short one-to-one conversations. They close with lessons about knowledge sharing and making sure important processes do not live only in one person's head.
Serving models is still the first problem many MLOps newcomers need to solve
Demetrios says 67 percent of the 1,000 people who completed the community's Campfire onboarding questionnaire selected model serving as the most important thing they wanted to learn. Monitoring, feature stores, databases, and data-centric AI did not rank as the main concern in that survey. Vishnu connects this to a common path into MLOps: a machine learning engineer builds a model, realizes the model needs to do something in production, finds there is no clear handoff, and starts learning how to serve it. He says this was also his own route into the community.
Community size matters less than repeated participation
Demetrios says the community grew from 2,600 people in January 2021 to about 7,500 by the end of the year. He calls the total membership a vanity metric because signups do not show whether people return or contribute. The more useful measure is active participation, which he puts at around 1,400 members. Vishnu says the growth in high-quality contributors and long discussion threads has been striking. He is also impressed that the quality of interaction has remained high as the community expanded.
The podcast became a way for its hosts to learn about production machine learning
Demetrios describes interviews with Cody Coleman and D. Sculley as his favorite podcast moments. Cody's conversation helped him understand how much data engineering, data warehouses, orchestration, and other parts of the modern data stack support machine learning in production. Interviewing D. Sculley felt like a personal milestone because Sculley had co-written the 2015 'High-Interest Credit Card' paper. Demetrios also recounts learning that Cody appeared in Angela Duckworth's book 'Grit', which connected a book that affected him personally with a guest he later interviewed.
MLOps jargon can hide how little a tool is used in a given setting
Demetrios is surprised that someone working on YouTube's recommendation system did not know what Kubeflow was. Vishnu uses that story to discuss how software and machine learning engineering often overemphasize fashionable tools, concepts, and languages. He cites Vicki Boykis's post 'You Don't Need Apache Kafka' as an example of questioning assumptions created by job postings and online discussion. Vishnu says engineers should focus on the problems they need to solve and the scale they actually operate at. They do not need to adopt every tool associated with large companies.
A sound analytics foundation comes before advanced machine learning
Vishnu says his career began with fitting models to CSV files, then moved through serving, training infrastructure, data engineering, analytics, and cloud DevOps. In 2022, he wants to understand the foundation better. He refers to 'The Informed Company', which describes a modern analytical stack as the foundation for an effective approach to machine learning. That foundation includes querying clean data in a warehouse, having a data model, and producing basic business metrics. Vishnu says a team that cannot tell whether its data is clean or produce basic business intelligence visualizations may not be ready to apply machine learning in its business.
The ecosystem's growth makes problem-first thinking more useful
Demetrios lists European AI regulation, KServe's graduation from the Kubeflow ecosystem, ClearML reaching version 1.0, and Teachable Hub reaching version 1.0. He also runs through acquisitions, funding rounds, company spinouts, and events such as MLOps World, TransformX by Scale AI, and Apply. Vishnu says the volume of funded companies, talks, blogs, and shared practices can feel overwhelming, even to someone running an MLOps podcast. His advice is to stay focused on the problem being solved and then find an appropriate tool, instead of trying to learn or use the whole ecosystem.
Community projects can turn existing engineering knowledge into practical learning
Vishnu praises the community's system design review project, where David studied a blog post about Pinterest's image similarity search system, interviewed its author, and turned the material into an animated video. Vishnu says the project showed the value of studying the process behind a finished engineering system. Demetrios also describes plans for a more engineering-focused community blog and a post arguing that the 'modern ML stack' is not a fixed thing. The hosts invite members to propose projects, write, host sessions, and build their own platform inside the community.
Knowledge sharing protects machine learning work from becoming a single-person dependency
Demetrios says an engineering manager at LaunchDarkly told him that engineers who share knowledge effectively tend to progress through their careers faster. He connects that lesson to a recurring podcast question: what happens if the person who owns a machine learning process suddenly becomes unavailable? The problem is especially difficult in MLOps because procedures and context can remain inside one person's head. Demetrios wants teams to create knowledge hubs and processes that allow someone else to perform the work. The goal is continuity when a person leaves, takes parental leave, or cannot do the job.
"How can you effectively make sure that there are the proper knowledge hub in place or the proper processes in place so that anyone can do it?"Demetrios Brinkmann49:30
Who should watch
You are trying to understand which MLOps problems teams actually face before choosing a tool.
You work in machine learning and need to build stronger data engineering or analytics foundations.
You help run an engineering community and want practical ideas for member onboarding, knowledge sharing, and contributor-led projects.
