What version are you on ?
Can you run a repair on the CF and check:
Does the repair detect differences in the CF and stream changes ?
After the streaming does it run a secondary index rebuild on the new sstable ? (Should be in the logs)
Can you provide the full query trace ?
Freelance Cassandra Consultant
TL;DR: I have inconsistend data (1 live row on node A & 1 tombstoned row on node B) that do not get fixed by repair. What can be a problem?
I have a CF containing Users' info, which I sometimes query by key, and sometimes by indexed columns like email. I'm using RF=2. I write with CL.ONE, but this CF is very rarely updated, so C* has a looot of time to fix inconsistencies that may occur, so I'm fine with this (at least in theory ;-) ).
To be clear:
- I've run a successfull cluster-wide repair on this CF before testing, so I do not expect any inconsistency
- All indexes are built, I've rebuilt them manually before testing, so I expect them to work properly (I mention it because it seems to be somehow related to indexes, but I'm not sure - see below)
The problem is:
When I query (cqlsh) some rows by key (CL is default = ONE) I _always_ get a correct result. However, when I query it by indexed column, it returns nothing.
When tracing a query with CL.ALL in cqlsh, I get info that C* has:
Read 0 live cells and 1 tombstoned // for first replica node
Read 1 live cells and 0 tombstoned // for second replica node
When CL is ONE it's never asking second replica for data (possibly due to DynamicSnitch scores or so), so it returns nothing.
Switching to CL >= TWO obviously fixes this problem for us, but it's not the solution I'd like to use as I'd rather rely on fast read/write requests with CL.ONE + frequent repairs, allowing some short-term inconsistency.
Any ideas why it may happen that data are still inconsistent after repair? Is there something I could have missed?
I'm mainly surprised that repair does not fix this inconsistency in ANY way - either by pulling missing data to first replica _OR_ tombstoning it on second replica. First one would be correct (delete was made a long time ago and then the row reappeared), but both could make sense, as both will make the data consistent. In this state it's definitely inconsistent and I don't understand it :-)