Description
The javadoc for setting up offline scans over RFiles (InputFormatBase.setScanOffline in 1.4 or InputFormatBase.setOfflineTableScan in 1.5) includes a nice little comment to the effect that if a "non-standard" directory is used for Accumulo in HDFS (read as, if the default value for instance.dfs.dir), accumulo-site.xml may need to be on the classpath for the mappers.
Best as I can tell, even if accumulo-site.xml is on the classpath, it makes no difference as InputFormatBase is creating a new ZooKeeperInstance which, in turn, will only ever make a DefaultConfiguration and never try to check if an accumulo-site.xml file is available. This would make it impossible for a non-default value for instance.dfs.dir to ever be used.