That's not what I saw in my test. I'm probably making some noob mistakes. Can someone enlighten me? Here's what I did: 1) Bring up a cluster with three servers cs1,2,3, with their initial token set to 'foo3', 'foo6', and 'foo9', respectively. ReplicationFactor is set to 2 on all 3. 2) Insert 9 columns with keys from 'foo1' to 'foo9', and flush. Now I have foo1,2,3,7,8,9 on cs1, foo1,2,3,4,5,6, on cs2, and foo4,5,6,7,8,9 on cs3. So far so good 3) Bring down cs3 and wipe out its data directory 4) Bring up cs3 5) run repair Keyspace1 on cs3, the flush At this point I expect to see cs3 getting its data back. But there's nothing in its data directory. I also tried getting all columns with ConsistencyLevel::ALL to see if that'll do a read pair. But still cs3's data directory is empty. What am I doing wrong?
This is 0.5.1 BTW. Thanks, - Jianing On Fri, Mar 26, 2010 at 6:12 PM, Rob Coli <rc...@digg.com> wrote: > On 3/26/10 5:57 PM, Jianing Hu wrote: >> >> In a cluster with ReplicationFactor> 1, if one server goes down, will >> new replicas be created on other servers to satisfy the set >> ReplicationFactor? > > Yes, via Anti-Entropy. > > http://wiki.apache.org/cassandra/AntiEntropy > http://wiki.apache.org/cassandra/ArchitectureAntiEntropy > > It's worth noting that "hot" keys are likely to be re-replicated by Read > Repair before Anti Entropy is triggered. > > http://wiki.apache.org/cassandra/ReadRepair > > =Rob > > > >
