Publishing the cluster state

The master node is the only node in a cluster that can make changes to the cluster state. The master node processes one batch of cluster state updates at a time, computing the required changes and publishing the updated cluster state to all the other nodes in the cluster. Each publication starts with the master broadcasting the updated cluster state to all nodes in the cluster. Each node responds with an acknowledgement but does not yet apply the newly-received state. Once the master has collected acknowledgements from enough master-eligible nodes, the new cluster state is said to be committed and the master broadcasts another message instructing nodes to apply the now-committed state. Each node receives this message, applies the updated state, and then sends a second acknowledgement back to the master.

The master allows a limited amount of time for each cluster state update to be completely published to all nodes. It is defined by the cluster.publish.timeout setting, which defaults to 30s, measured from the time the publication started. If this time is reached before the new cluster state is committed then the cluster state change is rejected and the master considers itself to have failed. It stands down and starts trying to elect a new master.

If the new cluster state is committed before cluster.publish.timeout has elapsed, the master node considers the change to have succeeded. It waits until the timeout has elapsed or until it has received acknowledgements that each node in the cluster has applied the updated state, and then starts processing and publishing the next cluster state update. If some acknowledgements have not been received (i.e. some nodes have not yet confirmed that they have applied the current update), these nodes are said to be lagging since their cluster states have fallen behind the master’s latest state. The master waits for the lagging nodes to catch up for a further time, cluster.follower_lag.timeout, which defaults to 90s. If a node has still not successfully applied the cluster state update within this time then it is considered to have failed and is removed from the cluster.

Cluster state updates are typically published as diffs to the previous cluster state, which reduces the time and network bandwidth needed to publish a cluster state update. For example, when updating the mappings for only a subset of the indices in the cluster state, only the updates for those indices need to be published to the nodes in the cluster, as long as those nodes have the previous cluster state. If a node is missing the previous cluster state, for example when rejoining a cluster, the master will publish the full cluster state to that node so that it can receive future updates as diffs.

Elasticsearch is a peer to peer based system, in which nodes communicate with one another directly. The high-throughput APIs (index, delete, search) do not normally interact with the master node. The responsibility of the master node is to maintain the global cluster state and reassign shards when nodes join or leave the cluster. Each time the cluster state is changed, the new state is published to all nodes in the cluster as described above.