Low Description
After applying the fix from CASSANDRA-10299 to the cluster we started having a problem of ~20k small sstables appearing for the table with static data when running incremental repair.
In the logs there were several messages about flushes for that table, one for each repaired range. The flushed sstables were 0.000kb in size with < 100 ops in each. When checking cfstats there were several writes to that table, even though we were only reading from it and read repair did not repair anything.
After digging around in the codebase I noticed that defragmentation of data can occur while reading, depending on the query and some other conditions. This causes the read data to be inserted again to have it in a more recent sstable, which can be a problem if that data was repaired using incremental repair. The defragmentation is done in CollationController.java.
I guess this wasn't a problem with full repairs since I assume that the digest should be the same even if you have two copies of the same data. But with incremental repair this will most probably cause a mismatch between nodes if that data already was repaired, since the other nodes probably won't have that data in their unrepaired set.
------
I can add that the problems on our cluster was probably due to the fact that CASSANDRA-10299 caused the same data to be streamed multiple times and ending up in several sstables. One of the conditions for the defragmentation is that the number of sstables read during a read request have to be more than the minimum number of sstables needed for a compaction(> 4 in our case). So normally I don't think this would cause ~20k sstables to appear, we probably hit an extreme.
One workaround for this is to use another compaction strategy than STCS(it seems to be the only affected strategy, atleast in 2.1), but the solution might be to either make defragmentation configurable per table or avoid reinserting the data if any of the sstables involved in the read are repaired.