I've been experiencing a behavior that is undesirable and it seems like a bug that causes a high amount of wasted work.
I have a CF where all columns have a TTL, are generally all inserted in a very short period of time (less than a second) and are never over-written or explicitly deleted. Eventually one node will run a compaction and remove rows containing only tombstones greater than gc_grace_seconds old which is expected. The problem comes up when a repair is run. During the repair the other nodes that haven't run a compaction and still have the tombstoned rows "fix" the inconsistency and stream the rows (which contain only a tombstone which is more than gc_grace_seconds old) back to the node which had compacted that row away. This ends up occurring over and over and uses a lot of time, storage, and bandwidth to keep repairing rows that are intentionally missing. I think the issue stems from the behavior of compaction of TTL rows and repair. The compaction of TTL rows is a node-local event which will eventually cause tombstoned rows to disappear from the one node doing the compaction and then get "repaired" from replicas later. I guess this could happen for rows which are explicitly deleted as well. Is this a feature or a bug? How can I avoid repair of rows that were correctly removed via compaction from one node but not from replicas just because compactions run independently on each node? Every repair ends up streaming tens of gigabytes of "missing" rows to and from replicas. Cassandra 1.1.5 with size tiered compaction strategy and RF=3 -Bryan
