Just created a simple test that puts 5 different columns into HBase at 3 different ts, so you end up with 3 versions of 5 values in memcache + 2 storeFiles and compared the timing for the old and the new implementations.
The query was asking for 3 versions of 2 of the columns, so a fairly simple setup, with no deletes and stuff, which would make the difference even bigger. Timing old = 2517596 ns Timing new = 373170 ns old/new = 6.7x This test is only done on the server side from the regionserver so no contact with the client so far, but it is looking good. Worth mentioning is also that this improvement is compared to the implementation already using HFile.
