Cassandra 'compacts' the way Lucene 'merges' segments. One interesting new feature built into Lucene is [1] which avoids loading the source files into the system IO cache on compaction / merge.
Perhaps Cassandra already has this feature? 1. https://builds.apache.org/job/Lucene-trunk/javadoc/all/org/apache/lucene/store/DirectIOLinuxDirectory.html
