Avinash pointed out two bugs in my remove code. One is easy to fix, the other is tougher.
The easy one is that my code removes tombstones (deletion markers) at the ColumnFamilyStore level, so when CassandraServer does read repair it will not know about the tombstones and they will not be replicated correctly. This can be fixed by simply moving the removeDeleted call up to just before CassandraServer's final return-to-client. The hard one is that tombstones are problematic on GC (that is, major compaction of SSTables, to use the Bigtable paper terminology). One failure scenario: Node A, B, and C replicate some data. C goes down. The data is deleted. A and B delete it and later GC it. C comes back up. C now has the only copy of the data so on read repair the stale data will be sent to A and B. A solution: pick a number N such that we are confident that no node will be down (and catch up on hinted handoffs) for longer than N days. (Default value: 10?) Then, no node may GC tombstones before N days have elapsed. Also, after N days, tombstones will no longer be read repaired. (This prevents a node which has not yet GC'd from sending a new tombstone copy to a node that has already GC'd.) Implementation detail: we'll need to add a 32-bit "time of tombstone" to ColumnFamily and SuperColumn. (For Column we can stick it in the byte[] value, since we already have an unambiguous way to know if the Column is in a deleted state.) We only need 32 bits since the time frame here is sufficiently granular that we don't need ms. Also, we will use the system clock for these values, not the client timestamp, since we don't know what the source of the client timestamps is. Admittedly this is suboptimal compared to being able to GC immediately but it has the virtue of being (a) easily implemented, (b) with no extra components such as a coordination protocol, and (c) better than not GCing tombstones at all (the other easy way to ensure correctness). Thoughts? -Jonathan
