There is no other way to make the cluster "forget" a node w/o decommission / removetoken.
You could do everything up to "stop the entire cluster" and do a rolling restart instead, kill the 2 nodes you want to remove, and then do removetoken, which would still do extra i/o but at least the slow nodes would not be involved. On Thu, May 20, 2010 at 8:54 PM, Ran Tavory <ran...@gmail.com> wrote: > I forgot to mention that the cluster cannot be taken down, it needs to > continue serving... is there another way? > > On May 21, 2010 3:03 AM, "Jonathan Ellis" <jbel...@gmail.com> wrote: > > One possibility: > > rsync the data to the next node in the ring that is in the same DC. > (specifically, rsync once, then flush on the source node and rsync > again.) Then stop the entire cluster, and restart everyone but those > two nodes. Then run nodetool repair on each machine. > > If your client is not reading at CL.ALL during repair, it could miss > data that was written after the rsync. Your call if that's > acceptable. > > On Wed, May 19, 2010 at 11:57 PM, Ran Tavory <ran...@gmail.com> wrote: >> In my cluster setup I have... > > -- > Jonathan Ellis > Project Chair, Apache Cassandra > co-founder of Riptano, the source for professional Cassandra support > http://riptano.com > -- Jonathan Ellis Project Chair, Apache Cassandra co-founder of Riptano, the source for professional Cassandra support http://riptano.com
