Our examples throughout this book have built, trained, and tested models only to destroy them a millisecond later. We can get away with this because our examples use limited training data and, at worst, take only a few minutes to train. Production applications will typically use much more data and require more time to train. In production applications, the trained model itself is a valuable asset that should be stored, saved, and loaded on demand. In other words, our models must be serializable.
Serialization itself is typically not a difficult issue. Models are essentially a compressed version of the training data. Some models can indeed be very large, but they will still be a fraction of the size of the data that trained them. What makes the topic of serialization challenging is that it opens up many other architectural questions that you will have to consider...