rocksdb: 1-write n-read, engine level distribution

[ ] sync in-memory change
[ ] load sst after compaction/flush
[ ] since writer/master may delete file if not needed any more, so it must "know" files in using by reader/slave
- explicit communication between writer & reader is required
- tailing on wal-log is simpler & slower than ZeroMQ
- manifest files also need to be monitored
[ ] there are two kind of readonly instance
- readonly & sync writer's update on DB, used for OLTP
- readonly & do not sync writer's update, used for OLAP(spark...), and distributed compaction

krareT / pub-task