Elasticsearch clustering

The Elasticsearch service can be deployed either on a single host in a "standalone" mode, or in a cluster of multiple hosts. The cluster mode will be enabled automatically after a few important variables and inventory groups are configured.

Standalone mode vs cluster mode

In a standalone mode, the Elasticsearch service will not try to talk with any other Elasticsearch nodes. Service will be usable over localhost connection. This mode is good for prototyping, testing and development environments, however it's not very resilient.

In a cluster mode, multiple Elasticsearch nodes talk to each other in a configured network subnet, over TCP. Elasticsearch clients communicate with the cluster over HTTP REST interface, usually via a dedicated host with Kibana and/or Logstash as an intermediary.

Playbook execution

When multiple Elasticsearch hosts are managed as a cluster, any changes in the cluster configuration should be implemented on all hosts in the cluster at the same time to avoid issues with split-brain or quorum. The role uses inventory groups to compute some specific values for all hosts in the cluster, however using the --limit parameter of the ansible-playbook command will only configure those values on a subset of hosts. Remember to always keep the whole cluster configuration synchronized by running the Elasticsearch playbook on all hosts included in the cluster (without the --limit parameter).

Ansible inventory groups

The debops.elasticsearch role uses a set of Ansible inventory groups to detect the Elasticsearch node type and change the configuration accordingly.

The main inventory group is [debops_service_elasticsearch]. Hosts in this group are configured to behave in the same way - all of them are eligible to be a master host, all of them can hold data, and all of them can use an ingest pipeline to process the input. This group is useful in small clusters, typically <10 hosts in total.

In larger clusters, the system administrator may want to separate the cluster hosts into separate node types. Each Ansible inventory group enables a separate feature, and hosts can be in multiple groups at once to mix and match the desired features:

Hosts in this Ansible inventory group are eligible to become masters.
Hosts in this Ansible inventory group can hold data shards.
Hosts in this Ansible inventory group can process incoming data via an ingest pipeline.
Hosts in this Ansible inventory group do not have any features explicitly enabled, and act as load balancers and coordinators within the cluster.

You can check the Elasticsearch node documentation for more details about node features.

The inventory groups and their corresponding node functions are defined using default variables. The role uses Ansible inventory groups to automatically determine the list of hosts which will be used for discovery, as well as the number of eligible master hosts, therefore direct changes to the node function variables should be done with care.

Unicast host discovery, number of master hosts

The role automatically manages the list of hosts which should be contacted for initial host discovery and number of master-eligible nodes based on the Ansible inventory group membership.

If the [debops_service_elasticsearch_master] group is not used, all of the hosts in the [debops_service_elasticsearch] inventory group will be added to the unicast discovery list, and all of them will be eligible to become masters.

When hosts are included in the [debops_service_elasticsearch_master] inventory group, only hosts in this group will be able to become masters, and only hosts in this group will be used for initial unicast discovery. Remember to always include an odd number of master-eligible hosts to achieve quorum majority within the cluster.

Firewall configuration

The role supports a firewall managed by the debops.ferm Ansible role. When the firewall is enabled, Elasticsearch will be configured to listen to connections on private IP addresses defined on the host along with the localhost; if the firewall is not detected or disabled, Elasticsearch will listen only on the localhost interface by default.

To enable cluster mode, you need to define at least one IP address or a CIDR subnet in the elasticsearch__allow_tcp list. Make sure to only allow access from trusted hosts!

There's also a separate elasticsearch__allow_http variable, but you don't need to enable it unless you need a direct access to the Elasticsearch HTTP REST interface from remote hosts. Kibana and Logstash installed on the same host as an Elasticsearch service should be able to talk to it over localhost with no issues.