MAPREDUCE-2264, a segment's raw data length is calculated without the EOF_MARKER bytes. However, when the merge is counting how many bytes it processed, it includes the marker. This can cause the merge progress to go above 100%.
Whether these EOF_MARKER bytes should count should be consistent between the two.
This a JIRA instead of an amendment because
MAPREDUCE-2264 already went into 2.0.3.