> When I say production machine learning, I’m referring to machine learning that manifests as a product feature. For example, Uber’s ETA prediction, or Gmail’s

> When I say production machine learning, I’m referring to machine learning that manifests as a product feature. For example, Uber’s ETA prediction, or Gmail’s Smart Compose. You can bet that prod services from companies you heard of are running on something more analogous to versioned docker images. Not a yaml file which says, 'Go run whatever predict.py is in the current folder.' The moment one of your dependencies breaks your code, or snookers your performance, there will be a lot of head scratching going on.