Cassandra repairs missing or inconsistent data in one of two ways: through constantly running background process of read repair or through the nodetool repair command.
The first method happens continuously if the consistency level requested was low. In this case, a background thread sends out the newest version of the data to those replicas which are out of data.
The second method, via nodetool repair, is an expensive operation in both disk and CPU consumption. The nodetool repair command uses a process called Anti-Entropy to build a Merkle tree from the data on the target node which will be compared with other nodes to discover inconsistencies. To monitor the progress of a nodetool repair operation, you can watch the active task queue in CompactionManager in conjunction with the active and pending counts on AE-SERVICE-STAGE and STREAMING-STAGE (which should be zero).