As Cassandra's log-based storage is timestamp-driven, it is absolutely critical that all nodes in a cluster have synchronized clocks. To this end, instances running Cassandra should be using NTPD or another time synchronization application to keep all system clocks in a cluster in sync.
Issues related to time synchronization discrepancies can manifest themselves as obsolete data or deleted data ghosting itself.
Consider the following current times from instances in a three-node cluster:
Let's say I run the following CQL to create a simple table: ...