Description
Split locations - at least for FileInputSplits - are generally tied to the location on HDFS where the split resides.
There are situations in which this location is not necessarily the best location to process this split.
e.g.
Clusters where compute and storage are separate.
Systems which cache data - cache affinity is more important.
Providing an alternate list of preferred locations allows grouping to the preferred locations, instead of always grouping based on the locations specified in the split.