During compaction range tombstones are not merged, causing performance degradation across the cluster.
The problem manifests in data partitions growing too large, and SSTables containing entries with duplicate range tombstones. This situation leads to overall growth in the number of rows within SSTables, as well as additional SSTable creation.
Read performance can degrade with a higher number of tombstones, causing increased JVM garbage collection activity, affecting overall cluster performance.
Typically if you look at the content of the sstable using sstable2json you’ll see a duplication such as:
The following jira outlines this issue: CASSANDRA-7953
Upgrade to a DSE release which contains the fix for the above Cassandra jira. Check the release notes for the jira or a minimum version of Cassandra 2.1.12 or later.