Details
-
New Feature
-
Status: Open
-
Minor
-
Resolution: Unresolved
-
None
-
None
-
None
Description
MapReduce provides the counters "org.apache.hadoop.mapreduce.JobCounter DATA_LOCAL_MAPS" and "org.apache.hadoop.mapreduce.JobCounter RACK_LOCAL_MAPS" that count how many local map tasks were started. This works great, if we have one HDFS block per a input split.
If we have multiple HDFS blocks in an input split (e.g. non-splittable file such as gzip, or use of CombineFileInputFormat), then these counters become inaccurate.
If it has not been proposed yet, I propose adding new counters that count how many bytes of data were read from a local node and a local rack. Thanks to them, a user will see the real data locality ratio.