After seeing the I/O issues in
CASSANDRA-1470, I've been doing some more thinking on this subject that I wanted to lay out.
I propose we redo the concept of how compaction works in Cassandra. At the moment, compaction is kicked off based on a write access pattern, not read access pattern. In most cases, you want the opposite. You want to be able to track how well each SSTable is performing in the system. If we were to keep statistics in-memory of each SSTable, prioritize them based on most accessed, and bloom filter hit/miss ratios, we could intelligently group sstables that are being read most often and schedule them for compaction. We could also schedule lower priority maintenance on SSTable's not often accessed.
I also propose we limit the size of each SSTable to a fix sized, that gives us the ability to better utilize our bloom filters in a predictable manner. At the moment after a certain size, the bloom filters become less reliable. This would also allow us to group data most accessed. Currently the size of an SSTable can grow to a point where large portions of the data might not actually be accessed as often.