This document provides an architect's view of Hadoop I/O based on analysis using vProbes instrumentation. It summarizes the results of a controlled small-scale study on a single-node Hadoop cluster running TeraSort. The study found that mapper tasks generate multiple temporary spill files and the reducer performs a large volume of shuffle I/O. It also presents initial observations about the Hadoop I/O model, including that mapper spill files account for 75% of disk bandwidth and HDFS input/output accounts for 12% of total bandwidth.