Part 3 Developing production machine learning code

Once a project is ready to ship to production, a few final tasks remain before the implementation is ready to be scheduled for deployment. While it’s tempting to see a test-passing build of an implementation as a complete and ready-for-the-real-world deployment, a few items need to be considered to ensure that people on call aren’t getting paged every few hours.

From drift monitoring, to principles of code architecture (which will aid in performing final peer reviews), prediction-quality assurance, logging, and serving infrastructure, these last items are the most oft overlooked. When ignored, they are some of the most regrettable elements to forget for those who have lived without their proper design and implementation.

In this section, we’ll go over these more advanced topics that can help make production deployment easier and help ensure that your models are explainable, able to be retrained, monitored, and (relatively) easy to update.