What Makes Spatial Data Big? A Discussion on How to Partition Spatial Data

摘要

The amount of available spatial data has significantly increased in the last years so that traditional analysis tools have become inappropriate to effectively manage them. Therefore, many attempts have been made in order to define extensions of existing MapReduce tools, such as Hadoop or Spark, with spatial capabilities in terms of data types and algorithms. Such extensions are mainly based on the partitioning techniques implemented for textual data where the dimension is given in terms of the number of occupied bytes. However, spatial data are characterized by other features which describe their dimension, such as the number of vertices or the MBR size of geometries, which greatly affect the performance of operations, like the spatial join, during data analysis. The result is that the use of traditional partitioning techniques prevents to completely exploit the benefit of the parallel execution provided by a MapReduce environment. This paper extensively analyses the problem considering the spatial join operation as use case, performing both a theoretical and an experimental analysis for it. Moreover, it provides a solution based on a different partitioning technique, which splits complex or extensive geometries. Finally, we validate the proposed solution by means of some experiments on synthetic and real datasets.