Sunday, January 31, 2016

Quorum and minimum_master_nodes setting for elasticsearch configuration

Theory behind elasticsearch recovery

Elastic search recovery works by having a master election satifying a particular minimum_master_nodes criteria. That is a master node will only be elected if there is a N number of master-eligible nodes (nodes in which node.master=true in their elasticsearch.yml file) to join it. N is specified by discovery.zen.minimum_master_nodes in the elasticsearch.yml file.

For  the quorum scheme, N should be set to

discovery.zen.minimum_master_nodes = (number of master-eligible nodes) / 2 + 1

The recovery works like this. When the current master node dies, a new master node will only be elected if there is N master-eligible nodes to join it, where N = (number of master-eligible nodes) / 2 + 1

Once the new master node is elected, if later the originally dead master node comes alive. It will have less than N master-eligible nodes to join it, therefore it will have to step down. Thus this scheme ensures that there will no two masters at the same time (which is the so-called split-brain scenario, that is not desirable since all master nodes have higher authority in its being able to update cluster-state in the data nodes and client nodes)

Minimum number of master eligible nodes required for an elasticsearch cluster

The minimum number of master-eligible nodes should be 3. And the zen.discovery.minimum_master_nodes should be equal = 3 / 2 + 1 = 2.

Reason: Suppose we only have two master-eligible nodes. If the master node dies, there is only 1 master-eligible node left, two things will happen depending on the value in discovery.zen.minimum_master_nodes:

Case 1:  if the zen.discovery.minimum_master_nodes is set to greater than 1, then there won't be any new master node elected, and the cluster will not operate.


Case 2:  if we set the zen.discovery.minimum_master_nodes=1, the new master node will be elected, however, when the originally dead master is brought alive again, the original master will not step down since it now also has one master-eligible node to join it, leading to split-brain problem.

Therefore the recommended minimum settings is to increases the number of master-eligible nodes to 3, and set the zen.discovery.minimum_master_nodes=2.

No comments:

Post a Comment