Currently, our memstore flush algorithm is pretty trivial. We let it grow to a flushsize and flush a region or grow to a certain log count and then flush everything below a seqid. In certain situations, we can get big wins from being more intelligent with our memstore flush algorithm. I suggest we look into algorithms to intelligently handle HLog compactions. By compaction, I mean replacing existing HLogs with new HLogs created using the contents of a memstore snapshot. Situations where we can get huge wins:
1. In the incrementColumnValue case, N HLog entries often correspond to a single memstore entry. Although we may have large HLog files, our memstore could be relatively small.
2. If we have a hot region, the majority of the HLog consists of that one region and other region edits would be minuscule.
In both cases, we are forced to flush a bunch of very small stores. Its really hard for a compaction algorithm to be efficient when it has no guarantees of the approximate size of a new StoreFile, so it currently does unconditional, inefficient compactions. Additionally, compactions & flushes suck because they invalidate cache entries: be it memstore or LRUcache. If we can limit flushes to cases where we will have significant HFile output on a per-Store basis, we can get improved performance, stability, and reduced failover time.