MLOps is the practice of building and maintaining production machine learning systems. It's new, and it's not all going well.

In a recent newsletter Laszlo Sragner wrote about what is wrong on the nascent field. It's a bit of a rant, and I liked it.

ML is still in its infancy and is plagued by hype. Lots of people are trying to sell complete 'solutions' to enterprise customers. Often these are bound to fail. Turn-key, fully repeatable solutions do not exist yet.

People hear about a jazzy new algorithm or some big data solution that Google has used and try and shoehorn it into the enterprise project they are working on that has 1,000 data-points. Following fashion and fads is easier than thinking deeply about the costs and benefits of various approaches.

When deciding what algorithm/database/platform/stack to use, don't base your decision on what you read about in a research paper last week, base it on the parameters of your problem. If you don't know what the parameters of your problem are yet, spend some time working them out first.

Most of the problems of ML are just engineering problems; we have methods and approaches, developed over decades, that solve those. There are new problems in MLOps but many have simple solutions. It's good to hear the voice of reason.

🛎️ Why this matters: MLOps is certainly important but beware of the hype mongers and use common sense.