MLOps World Takeaways
Updated: 7 days ago
I had the privilege of attending MLOps World 2022 conference in Toronto earlier this month and it was incredibly rewarding to meet so many companies and thought-leaders from our industry. You tend to forget that underneath all the websites, branding, and social media personas that we are just a bunch of (relatively) normal humans!
2019 the Magic Year
A very interesting observation that my business partner Zev and I made at the conference was that a lot of MLOps companies seem to have been founded in 2019. This is completely anecdotal, but the date just kept coming up in conversations. Very few MLOps companies have been around for longer than 5 years which indicates how young this industry really is. What makes 2019 so special is hard to say, but there must have been some inflection point where enough companies were deploying Machine Learning models to warrant a need for standardization and outsourcing.
Oh We’re Both Doing this?
With the recent surge in independent MLOps tools - from model versioning, to deployment, to monitoring (and everything in between) - the overall package from the customer perspective has become increasingly fractured, and frankly, quite confusing. This is natural with the early formation of an industry and I think we are already starting to trend towards an equilibrium. The pandemic has likely exacerbated this by eliminating in-person events like trade shows, which provide venues for competing companies to see each other’s products and business models. Losing these opportunities makes it difficult for each company to “find their lane” in the industry where they can specialize in an underserved area that potentially compliments their competitors. At the MLOps World conference there was a sense of redundancy and lack of distinction across competing platforms and I suspect many leadership teams are now reflecting on what makes them unique.
Less Dashboards Please!
We heard grumblings on more than one occasion from end-users that there are simply too many product dashboards focusing on specific pieces of the MLOps pipeline. The MLOps movement is a push for standardization in developing, testing and deploying Machine Learning at industrial scale. As with any industrial process it is desirable to employ a cohesive environment across the entire process where information can easily flow between departments. In the context of Machine Learning, this means the data engineering, modeling, and Ops teams are all speaking a common language, even though they each focus on a specific part of the process.
There are a number of platforms offering a siloed environment focusing on their area of expertise, rather than working with upstream and/or downstream tools to provide a more integrated experience for the end user. MLOps teams have to navigate different platform environments - complete with unique logins, dashboard interfaces, user accounts, and metrics - as their models work their way through the pipeline. Cooperation amongst MLOps tools to bring their functionality under a common dashboard environment (a “single pane of glass”) will reduce barriers to adoption and present a more rewarding customer experience. Of course there will always be a need for isolated deep dive tools, but I think these will become the exception rather than the norm.
Selling Bikes or Handlebars
I’m a fan of analogies so let’s compare the state of the MLOps industry with selling bikes. We can all agree that the vast majority of people buying bikes want to purchase a complete bike. They don’t want to individually source a frame, wheels, handlebars, pedals, etc. and build their own bike. They just need something functional for transportation that’s easy to acquire and ready to go right out of the box. A small fraction of the market are avid cyclists and they do want to source their own components because they have very specific requirements that must be satisfied. This means as a handlebar manufacturer, most of your product is sold to end-users indirectly through a bike manufacturer who puts all the pieces together and sells the final product to the customer. A small portion of your product may be purchased directly from cyclists who specifically sought out your handlebars.
Right now I believe there’s a disconnect in the MLOps industry between how the tools are being sold vs. how the majority of customers want to acquire them. The individual MLOps tools (i.e. the bike parts) are being pushed on the end-users who mostly just want a complete product (i.e. a bike). This is presenting a barrier for the general-purpose customers who want a complete MLOps platform that is fully-assembled and easy to implement. It bodes well for the specialist customer (i.e. the avid cyclist) who does their research and knows exactly what they require, but they are a small representation of the overall demand. For the entire MLOps market to be realized, complimentary tools need to take a cooperative approach so they can provide a complete package.