Ask HN: How do you decide to release a new ML model in production?
I am an ML engineer and have my own workflow that I use now and an aspirational workflow that I want to use. But I was curious how other ML engineers make this call. Here are some options to kickstart your thinking: a) Release it if it performs well in offline training b) A/B test in production alongside existing model(s) c) Use contextual bandits in production d) Some other way (details appreciated :-) )
No comments yet.