Hello,
We have just moved from solr 4.6 master/slave to 6.4.2 SolrCloud. We have
three collections, each with a single shard and a varying number of replicas,
all kept by an ensemble of three zooKeepers (on their own hosts). As an
ecommerce site, our capacity needs vary so we add and remove replicas with some
frequency. The basic topology is like this:
solr1
|- collection1
|- shard1 - replica1
|- collection2
|- shard1 - replica1
|- collection3
|- shard1 - replica1
.
.
.
solrN
|- collection1
|- shard1 - replicaN
|- collection2
|- shard1 - replicaN
|- collection3
|- shard1 - replicaN
Where N varies between three and six most of the time.
During a recent test, we ran our indexing processes to a set of nodes, and then
two nodes were removed from our configuration. Subsequently the remaining
nodes were reindexed, without problems. The two nodes that had been previously
removed (by simply stopping solr on those boxes) were brought back into the
cluster by starting solr with the appropriate zkHost strings. (These were the
same zkHosts as when the instances were stopped.) We found that the indexes
did not synch up until we re-indexed the entire cluster.
What are we missing? We need the re-added indexes to synchronize with those
already active in the cluster. If we have to re-index the whole cluster, we
risk inconsistent results being served from the new nodes while indexing is
going on. In reviewing the Reference Guide and doing various searches, I
haven't found anything that clearly references adding replicas to a cluster
when the cores already contain data.
Thank you for any insights,
Joe
Joe Heasly, Systems Analyst I
L.L.Bean, Inc. ~ Direct Channel Business & Technology Team
Office: 207.552.2254
Cell: 207.756.9250
