2 Conceptual overview #

Each site runs one booth instance that manages communication with the other sites. If you have a setup with an even number of sites, it is useful to have an additional instance to reach consensus about decisions such as failover of resources across sites. In this case, add one or more arbitrators running at additional sites. Arbitrators are single machines that run a booth instance in a special mode. As all booth instances communicate with each other, arbitrators help to make more reliable decisions about granting or revoking tickets. Arbitrators cannot hold any tickets.

An arbitrator is especially important for a two-site scenario: for example, if site A can no longer communicate with site B, there are two possible causes for that:

However, if site C (the arbitrator) can still communicate with site B, site B must still be up and running.

Booth is the instance managing the ticket distribution, and thus, the failover process between the sites of a Geo cluster. Each of the participating clusters and arbitrators runs a service, the boothd. It connects to the booth daemons running at the other sites and exchanges connectivity details. After a ticket has been granted to a site, the booth mechanism can manage the ticket automatically. If the site that holds the ticket is out of service, the booth daemons will vote which of the other sites will get the ticket. To protect against brief connection failures, sites that lose the vote (either explicitly or implicitly by being disconnected from the voting body) need to relinquish the ticket after a time-out. Thus, it is made sure that a ticket will only be redistributed after it has been relinquished by the previous site. See also Dead man dependency (loss-policy="fence").

For a Geo cluster with two sites and arbitrator, you need 3 booth instances: one instance per site plus the instance running on the arbitrator.

Anmerkung: Limited number of booth instances

The upper limit is (currently) 16 booth instances.

After a ticket is revoked, it can take a long time until all resources depending on that ticket are stopped, especially in case of cascaded resources. To cut that process short, the cluster administrator can configure a loss-policy (together with the ticket dependencies) for the case that a ticket gets revoked from a site. If the loss-policy is set to fence, the nodes that are hosting dependent resources are fenced.

Warnung: Potential loss of data

On the one hand, loss-policy="fence" considerably speeds up the recovery process of the cluster and makes sure that resources can be migrated more quickly.

On the other hand, it can lead to loss of all unwritten data, such as:

• Data lying on shared storage.

• Data in a replicating database (for example, MariaDB or PostgreSQL) or on a replicating device (DRBD), where the data has not yet reached the other site because of a slow network link.

A ticket grants the right to run certain resources on a specific cluster site. A ticket can only be owned by one site at a time. Initially, none of the sites has a ticket—each ticket must be granted once by the cluster administrator. After that, tickets are managed by the booth for automatic failover of resources. But administrators may also intervene and grant or revoke tickets manually.

After a ticket is administratively revoked, it is not managed by booth anymore. For booth to start managing the ticket again, the ticket must be again granted to a site.

Resources can be bound to a certain ticket by dependencies. Only if the defined ticket is available at a site, the respective resources are started. Vice versa, if the ticket is removed, the resources depending on that ticket are automatically stopped.

The presence or absence of tickets for a site is stored in the CIB as a cluster status. With regard to a certain ticket, there are only two states for a site: true (the site has the ticket) or false (the site does not have the ticket). The absence of a certain ticket (during the initial state of the Geo cluster) is not treated differently from the situation after the ticket has been revoked. Both are reflected by the value false.

A ticket within an overlay cluster is similar to a resource in a traditional cluster. But in contrast to traditional clusters, tickets are the only type of resource in an overlay cluster. They are primitive resources that do not need to be configured or cloned.

After you have initially granted an automatic ticket to a site, booth manages this ticket automatically. If the site holding a ticket should be out of service, the ticket is automatically revoked after the expiry time. If the remaining sites have quorum, the ticket is granted to another site (fail over). The resources that depend on that ticket fail over to the new site that holds the ticket. The loss-policy (which is defined within the constraint) specifies what happens to the nodes that have run the resources before.

If automatic failover in case of a split-brain scenario is not required, administrators can also grant manual tickets to the healthy site. How to manage tickets from command line is described in Abschnitt 8.1.