Designing Machine Learning Systems

Chapter 8. Data Distribution Shifts and Monitoring

Let’s start the chapter with a story I was told by an executive that many readers might be able to relate to. About two years ago, his company hired a consulting firm to develop an ML model to help them predict how many of each grocery item they’d need next week, so they could restock the items accordingly. The consulting firm took six months to develop the model. When the consulting firm handed the model over, his company deployed it and was very happy with its performance. They could finally boast to their investors that they were an AI-powered company.

However, a year later, their numbers went down. The demand for some items was consistently being overestimated, which caused the extra items to expire. At the same time, the demand for some items was consistently being underestimated, leading to lost sales.¹ Initially, his inventory team manually changed the model’s predictions to correct the patterns they noticed, but eventually, the model’s predictions had become so bad that they could no longer use it. They had three options: pay the same consulting firm an obscene amount of money to update the model, pay another consulting firm even more money because this firm would need time to get up to speed, or hire an in-house team to maintain the model onward.

His company learned the hard way an important lesson that the rest of the industry is also discovering: deploying a model isn’t the end of the process. A model’s performance ...