I understand what you are thinking daniel, but this approach has at least one big wrinkle. You would be introducing depencencies between compaction and replication.
The 'unique' idempotent records are required for cassandra to read repair properly. Therefore, if a compaction (or even a memtable flush) occurred, the system could no longer read repair the counters. Your strategy is closer to how bt/hbase handles accumulators, but it works because in that system there is a single consistient write log.
Here is a different approach to doing this with cassandra...
- use timestamps in the column uniqueness
- Don't try to use custom compaction. Instead, layer counter summarization on top as a periodic summarization job.
- make sure the summarizer does try to do it's job for a batch of counters until they are fully replicated and 'static' (no new increments will appear)
- write the 'summary' of a bunch of unique timestamps in a way that anyone summing the values know the existance of a summary means they should ignore individual values for the range (because it will take time for them to be deleted)