A while ago I said:
ideally we would parallelize within a single sstable (breaking out the deserialize / merge / write stages) but this is Hard.
It's hard, but for a lot of users (anyone where a single CF holds the bulk of the data) this is the only kind of optimization that will make a difference.
There are five stages: read, deserialize, merge, serialize, and write. We probably want to continue doing read+deserialize and serialize+write together, or you waste a lot copying to/from buffers.
So, what I would suggest is: one thread per input sstable doing read + deserialize (a row at a time). One thread merging corresponding rows from each input sstable. One thread doing serialize + writing the output. This should give us between 2x and 3x speedup (depending how much doing the merge on another thread than write saves us).
This will require roughly 2x the memory, to allow the reader threads to work ahead of the merge stage. (I.e. for each input sstable you will have up to one row in a queue waiting to be merged, and the reader thread working on the next.) Seems quite reasonable on that front.
Multithreaded compaction should be either on or off. It doesn't make sense to try to do things halfway (by doing the reads with a
threadpool whose size you can grow/shrink, for instance): we still have compaction threads tuned to low priority, by default, so the impact on the rest of the system won't be very different. Nor do we expect to have so many input sstables that we lose a lot in context switching between reader threads. (If this is a concern, we already have a tunable to limit the number of sstables merged at a time in a single CF.)
IMO it's acceptable to punt completely on rows that are larger than memory, and fall back to the old non-parallel code there. I don't see any sane way to parallelize large-row compactions.