On 12/19/2017 3:06 PM, Greg Roodt wrote: > Thanks for your reply Erick. > > This is what I'm doing at the moment with Solr 6.2 (I was mistaken, before > I said 6.1). > > 1. A new instance comes online > 2. Systemd starts solr with a custom start.sh script > 3. This script creates a core.properties file that looks like this: > ``` > name=blah > shard=shard1 > ``` > 4. Script starts solr via the jar. > ``` > java -DzkHost=....... -jar start.jar > ```
The way that we would expect this to be normally done is a little different. Adding a node to the cloud normally will NOT copy any indexes. You have basically tricked SolrCloud into adding the replica automatically by creating a core before Solr starts. SolrCloud incorporates the new core into the cluster according to the info that you have put in core.properties, notices that it has no index, and replicates it from the existing leader. Normally, what we would expect for adding a new node is this: * Run the service installer script on the new machine * Add a ZK_HOST variable to /etc/default/solr.in.sh * Use "service solr restart"to get Solr to join the cloud * Call the ADDREPLICA action on the Collections API The reason that your method works is that currently, the "truth" about the cluster is a mixture of what's in ZooKeeper and what's actually present on each Solr instance. There is an effort to change this so that ZooKeeper is the sole source of truth, and if a core is found that the ZK database doesn't know about, it won't be started, because it's not a known part of the cluster. If this goal is realized in a future version of Solr, then the method you're currently using is not going to work like it does at the moment. I do not know how much of this has been done, but I know that there have been people working on it. Thanks, Shawn
