> It seems incongruous to say things aren't tested at scale is what's > holding us back from 1.0. And to say that HFileV3 can go in and still > go 1.0. >
Let me elaborate. There are at least new RPC PB, PB structures in HFiles / hlogs, and zk, table locks, bucket cache, online merge, stochastic LB, hbase on windows, new data types, AM changes, favorite node assignment, dist log replay, and tons of MTTR changes that are not run in production so far. The amount of new stuff is scary. Yes, we and other groups are doing extensive testing, but I fear that if we prematurely call this 1.0, then we would be asking for trouble. HFile v3 on the other hand, can be baked in 0.96.x releases if time allows it. I am expecting that to also go a few minor releases, and not be committed just before 1.0 comes out. Enis
