Whenever a node becomes unresponsive or down for a certain period of time, all the write requests to that node will fail. Cassandra keeps a copy of the partitions that were supposed to be written to the unresponsive node within a local hints table. The hints are stored for a certain configurable period of time. When the downed node comes up and starts gossiping with the rest of the nodes in the cluster, the hints are replayed to the downed node, and data is back in sync between all the replica nodes.
The coordinator node will eventually evict hints if the downed node doesn't come back up within the eviction period. The coordinator also stops storing hints related to a downed node if it is down for a configurable period of time. ...
