It’s an exciting time for the MLOps ecosystem, and there’s no better place to be than in Toronto! The MLOps World Summit 2022 happened last week in Toronto and truly lived up to its promise of being the ultimate ML Operations & strategy conference & Expo. It saw a number of MLOps companies and practitioners, including our Chief Data Architect, Achint Thomas, gathered together to discuss the state of machine learning (ML) in production. There were a bunch of very interesting takeaways from the event, which Achint was kind enough to summarize for all of us Interestingly, what we noticed was that there were three recurring themes during the event. These themes managed to find their way in almost every talk and presentation at the event, and pretty much dominated all of our conversations with folks in the community, as well as friends of Scribble Data.
Dominant Themes at MLOps World 2022
ML spend should be justified by business outcomes
Deploying machine learning to solve business problems is a high-cost investment. Selecting an algorithm and building ML models are only parts of the puzzle. Reliable access to trustworthy data (both raw and enriched), knowledge and understanding of the data itself (annotations), being aware of changes and drift in source data, explaining why a model does what it does with the data it sees, defining the right metrics to serve as early warning systems for when things go wrong, alerting the right set of stakeholders, inferencing at scale, monitoring deployed models in production, gathering and using model outputs and feedback to constantly improve model performance, and deciding whether and where to use humans-in-the-loop are all equally important aspects. Making all of this work together like a well-oiled machine requires high-skill talent and bringing together the right set of tools. Today, businesses are starting to ask whether their investment in ML for the sake of ML is yielding the expected ROI. Most folks at the summit agreed that ML practitioners would be well advised to always stay close to the specific business problems being solved, rather than build ML systems for the sake of technology. Scribble Data agrees.
Know when to build vs. buy
The build vs. buy dilemma is real – and even more acute in the case of real-world ML. While it may be fun to glue together the latest, greatest, open-source ML tooling for data prep, feature generation, model building, and model serving, the talent maintenance costs start to add up. There is value in being able to offload parts of the ML stack to vendors who do it well. As the last decade saw compute and storage move to the cloud, the MLOps space is also seeing a number of options spring up to take over the complexity of developing, running, and maintaining the ML stack. But we’re still early, and it’s not settled that there is a clear preference to “buy” vs. “build”, yet. Part of this dilemma boils down to the fact that complex ML approaches may not be the best fit for all problems. To a hammer, everything looks like a nail.
Sufficient complexity exists for niche offerings
Given the breadth of tooling and expertise needed to drive impact when deploying machine learning, it is not surprising that separate niches have formed over the past few years. These niches are being filled by very smart people and companies doing very smart things. In attendance were a number of companies, both young and more mature, representing providers of tools, platforms, and services. The following is a wholly incomplete list of companies in this space who attended the summit, and who are worth keeping an eye on:
-
ML Infrastructure (Redis, AWS, IBM, UbiOps, Tenstorrent, Genesis Cloud, Qwak AI)
-
Feature stores (Tecton, Featureform)
-
MLOps platforms (VesslAI, Pentuum, Pachyderm, Iguazio, Modzy, ClearML, Seldon)
-
ML observability and explainability (Aporia, Superwise, Arize)
-
ML security and safety (TrojAI, Robust Intelligence)
-
ML-powered applications (PrivateAI)
-
ML talent (Sharpest Minds, Harnham)
The future of MLOps is exciting!
The conversations at MLOps World Summit only strengthened our belief in what we’re building – tools that can help organizations put data to work to drive business outcomes. If you’re in the business of data or wondering how you can leverage data for key decision-making, we’d love to hear from you.
And if you’re looking for your next adventure and would like to play a key role in how organizations can significantly reduce friction in the consumption of data, we’d love to talk to you about open roles we are hiring for. Check out https://www.scribbledata.io/careers and drop us a line!