We have a load balancing policy which selects the host best on latency and uses a Phi convict algorithm in a method similar to DynamicSnitch. Using this policy, you would inherently get the closest replica whenever possible as that would most likely be the best performing.
This policy is still in trunk and 0.7.0 tip. We should have a new release out containing the above in the next few days. On Thu, Mar 24, 2011 at 8:46 AM, Jonathan Colby <jonathan.co...@gmail.com> wrote: > Indeed I found the big flaw in my own logic. Even writing to the "local" > cassandra nodes does not guarantee where the replicas will end up. The > decision where to write the first replica is based on the token ring, which > is spread out on all nodes regardless of datacenter. right ? > > On Mar 24, 2011, at 2:02 PM, Jonathan Colby wrote: > >> Hi - >> >> Our cluster is spread between 2 datacenters. We have a straight-forward IP >> assignment so that OldNetworkTopology (rackinferring snitch) works well. >> We have cassandra clients written in Hector in each of those data centers. >> The Hector clients all have a list of all cassandra nodes across both data >> centers. RF=3. >> >> Is there an order as to which data center gets the first write? In other >> words, would (or can) the Hector client do its first write to the cassandra >> nodes in its own data center? >> >> It would be ideal it Hector chose the "local" cassandra nodes. That way, if >> one data center is unreachable, the Quorum of replicas in cassandra is still >> reached (because it was written to the working data center first). >> >> Otherwise, if the cassandra writes are really random from the Hector client >> point-of-view, a data center outage would result in a read failure for any >> data that has 2 replicas in the lost data center. >> >> Is anyone doing this? Is there a flaw in my logic? >> >> > >
