On 10/28/2015 7:21 AM, andreic9203 wrote: > Because in production we have a lack of hardware, our environment contains > the following: > 1 Zookeeper instance > 2 collections (collection1 and collection2) > 3 Sol'r nodes (node1, node2 and node3) > > collection1 is shared between node1 and node2 > collection2 is used by node3 > > On the central machine I have the zookeeper instance, node1 and node3 > instances. > I have another machine which contains the node2 instance. > > Between these 2 machines, the internet connection is pretty unstable. From > time to time, the connection is lost and the node2 become unavailable. > Sometimes after the connection is restored, the node2 connects successfully > to Zookeeper but sometimes it does not.
SolrCloud is not designed for this situation. Connectivity between Solr nodes and zookeeper must be solid. SolrCloud does not work well in a two-location geographically dispersed setup. For zookeeper redundancy, you need a minimum of three zookeeper instances in an ensemble, because at least two of them must be fully operational and reachable, or the ensemble loses quorum. This ensemble should be co-located with all Solr nodes that use it. You might be thinking you could put two zookeeper nodes in each location, but this doesn't work, because when there are four zookeeper nodes, at least three of them (a majority) are required to maintain quorum. This problem continues as you scale up ... if there are six nodes (three at each location), then four of them are required for quorum. A geographically dispersed zookeeper/solrcloud configuration requires three locations, not two, and if more than one of those datacenters loses connectivity, then you're not completely operational. Here's the detail in the zookeeper documentation that describes requirements for failure scenarios: http://zookeeper.apache.org/doc/r3.5.1-alpha/zookeeperAdmin.html#sc_CrossMachineRequirements There is a SolrCloud feature coming called Cross-Data-Center Replication, which will let you replicate between separate clouds in two data centers, each with its own local zookeeper ensemble. I do not know anything about when that feature will be available. Thanks, Shawn
