thanks Stack.
I only started looking into hbase after comparing it against Cassandra (I've been using Cassandra for a while) HBase enforces a single access point (the active region server) among (well, actually not "among") all replicas at any time; while cassandra allows client to access any of the replicas; among other things, one advantage of the HBase approach is that client achieves strong consistency while reading/writing only one node (the region server), while Cassandra needs to read/write multiple; on the flip side, HBase could see a short unavailability interim period for the region server switching, while Cassandra does not have such switching, so won't see such unavailability periods. Yang On Mon, Jul 11, 2011 at 9:15 AM, Stack <[email protected]> wrote: > On Mon, Jul 11, 2011 at 9:10 AM, Yang <[email protected]> wrote: >> if the current region server fails, and master detects that and >> assigns the region to another region server, >> the new region server needs to read the SSTables and HLog and >> construct the memtable. >> > > Yes. > >> the master reassignment+memtable construction process, how long >> roughly does it take? >> during this process, the region is unavailable to clients (though for >> only a short while), right? >> if so, I wonder how actual HBase installations in production handle this? >> > > It depends on how many logs need splitting. It can take seconds to > minutes. If lots of logs, many minutes. During this time the > region/data is offline for reads and writes. > > In 0.90.x hbase, the split process is done by the master process. In > 0.92.0, the log splitting is distributed so should run faster in > general. > > St.Ack > St.Ack >
