ProdOps - ML Products in Production
2021 Dec 25ProdOps - ML Products in Production
A conversation I had with Christiano Milfont and Matheus Dias about Data-Driven Culture is now on YouTube; but our talk was much more about Machine Learning in production.
We discussed a case I worked on back at MyHammer, where we had a 45-day deadline to deliver an ML API for our production request classification service using NLP. I talked a bit about the modeling in this MyHammer post at the time.
Personally, I was very proud of delivering this case because it was essentially less conventional than most cases we see on the internet, where the data science part usually comes at the beginning of the workflow:
The big difference in our mindset was that we treated our problem not as a Data Science problem where a model might potentially be put into production after a scientific process and all; but rather as a service integration case for an API that would deliver Function-as-a-Service to the backend and frontend teams.
The rationale behind the decision is detailed in the video; but in short, an “API First” approach helped us a lot in delivering the model to production from day one, giving us the confidence that when it was actually in production, we would have all the elements for monitoring, observability, and so on.
I’m still going to write a blog post about this ProdOps philosophy from an ML perspective, given that I personally think the discussion in our industry is still very focused on tools and ML itself, and not enough on how these technologies can be used in production at scale.
