Hey everyone, have some questions about clustering. Most of my problems
have been solved just reading this group, but I'm getting close to going
into production and had a couple questions.
Currently if I restart the master, the cluster breaks up, or at least the
nodes all declare themselves new masters.
For whatever reason multicast is disabled on our network. Does this mean
that each host must have a list of all other hosts in the cluster, like
this:
Hey everyone, have some questions about clustering. Most of my problems have been solved just reading this group, but I'm getting close to going into production and had a couple questions.
Currently if I restart the master, the cluster breaks up, or at least the nodes all declare themselves new masters.
For whatever reason multicast is disabled on our network. Does this mean that each host must have a list of all other hosts in the cluster, like this:
Currently if I restart the master, the cluster breaks up, or at least
the nodes all declare themselves new masters.
For whatever reason multicast is disabled on our network. Does this
mean that each host must have a list of all other hosts in the
cluster, like this:
Multicast/unicast discovery is used only when starting a new cluster,
not when a cluster is already running.
The setting that you want is minimum_master_nodes. This specifies the
minimum number of master-eligible nodes (ie nodes that are allowed to
become master) that a node must be able to see in order for it to form a
cluster. If a node sees to few other nodes, then it doesn't form a
cluster, but keeps looking for more nodes or an existing cluster.
minimum_master_nodes should be set to: int(num_nodes/2)+1 (ie the
majority)
What you describe above seems very weird to me. If you have managed to
form a cluster of eg 4 nodes, then shutdown the master, the other three
nodes should elect a new master and continue as a cluster. They
shouldn't disconnect from each other (unless of course they are unable
to communicate).
Try replicating this situation, and check the cluster health on all
nodes while you're doing it (and the logs on all nodes) to figure out
exactly what is happening.
Apache, Apache Lucene, Apache Hadoop, Hadoop, HDFS and the yellow elephant
logo are trademarks of the
Apache Software Foundation
in the United States and/or other countries.