MLSys is similar to traditional computer systems engineering in that they both involve performance, memory efficiency, and distributed computing. The key difference is the workload: machine learning. Because ML workloads behave differently from traditional tasks, the industry built a unique stack from hardware to software, distinct from the classic operating systems stack (see Figure 1). Due to its fullstack nature, MLSys is challenging, and by tackling it you will gain substantial technical depth.

This book focuses on the machine learning system layer. We will break MLSys into finer-grained components later in the book.