Spark GraphX is not the first system to bring graph parallel computation to the mainstream, as even before this project, people used to perform graph computation on Hadoop, although they had to spend considerable time to build a graph on Hadoop. This resulted in creation of specialized systems such as Apache Giraph (an open source version of Google's Pregel), which ensured that the graph processing times come down to a fraction of what they were on Hadoop. However, graph processing is not isolated, and is very similar to MLLib where you have to spend time to load the data and pre-process it before running a machine learning pipeline. Similarly, the full data processing pipeline isn't just about running a graph algorithm, and graph creation is an important aspect of the problem, including performing post-processing, that is, what to do with the result. This was beautifully presented in a UC Berkley AmpLab talk in 2013 by Joseph Gonzalez and Reynold Xin.
The following figure...
