ML System Stack

The amount of data that can be generated and stored in academic and industrial projects and applications is increasing rapidly. Big data analytics technologies have established themselves as a solution for big data challenges to the scalability problems of traditional database systems. The vast amounts of new data that is collected, however, usually is not as easily analyzed as curated, structured data in a data warehouse is. Typically, these data are noisy, of varying format and velocity, and need to be analyzed with techniques from statistics and machine learning rather than pure SQL-like aggregations and drill-downs. Moreover, the results of the analyses frequently are models that are used for decision making and prediction. The complete process of big data analysis is described as a pipeline, which includes data recording, cleaning, integration, modeling, and interpretation.

In this lecture, we will discuss big data systems, i.e., infrastructures that are used to handle all steps in typical big data processing pipelines. We will learn about data center infrastructure and scale-out software systems. The software discussed will cover the full big data stack, i.e., distributed file systems, Map Reduce, key value stores, stream processing, graph processing, ML systems.