7.3.1. Tungsten Manager Definitions

Content Being Written

This section of the documentation is currently being produced and may be incomplete and/or subject to change.

These definitions assume a familiarity with concepts like failover, switch, master and slave datasource etc.

  • coordinator - every cluster designates one of the Tungsten managers in the cluster as the coordinator and it is this manager that will be responsible for taking action, if action is required, to recover the cluster's database resources to the most highly available state possible.

  • rules - this term specifically refers to a set of 'business rules', implemented in a format required by the 'JBoss Drools' rules engine, and which are used to perform fault detection, fencing and recovery for Tungsten Clustering.

  • rule firing/triggering - refers to the action of a rule becoming active due to the coincidence of one or more conditions specified in the rule itself. For example, a rule that detects a potential missing cluster member, called a 'heartbeat gap detection' rule, can fire if there are no other active alarms and if the rule has not 'seen' a cluster member heartbeat within the last 30-45 seconds. If a rule 'fires', further processing, specified in Java, will take effect.

  • fault/fault detection - any condition which, if left unresolved, could lead to a lack of availability of database resources or to data inconsistency etc. Faults are detected. An example of a fault detection is to detect that a specific database server has stopped.

  • (fault) alarm - the Tungsten Manager uses raises and processes entities that we will refer to as 'alarms' as an initial part of fault detection. A set of manager rules raise alarms under specific circumstances and that alarm stays active or triggers further processing depending on other rules. An alarm, depending on its type, may not necessarily mean that a fault has been definitively detected but that something has been detected that may or may not lead to an actual fault condition. An example of this, as you will see later, is a HeartbeatGapAlarm which occurs when the rules on a specific manager detect that a heartbeat event has not been received from one or more of the other managers in the group.

  • (fault) alarm retraction - the action taken, by the rules, to remove an alarm from from consideration for further action. For example, if the manager raises a 'Heartbeat Gap Alarm' and then, subsequently, detects that the heartbeat from the errant member has resumed, a rule will retract that alarm.

  • (fault) fence/fencing - the action which leads to a first-level amelioration of a fault i.e. an action that keeps a fault from causing further harm - applications are isolated from the fault through the action of fencing. Fencing a fault results in the removal of the original fault condition but may also result in further recovery actions. An example of fencing the fault that is detected when a database server has stopped is to set the associated datasource to the FAILED state. This effectively makes the datasource unavailable to applications, immediately, thus isolating the application from the fault state.

  • (fault) recover/recovery - the action which may occur after a fault is initially fenced and which leads to a condition of continued availability, data consistency etc. An example of a recovery operation is, for example, the failover that occurs after a stopped master database fault has been fenced.

  • split-brain - a condition of a cluster such that members of the cluster group have different views of the same set of database resources and, in the most damaging incidence of split-brain, more than different cluster members designate different database resources as the master resource, resulting in applications being able, for example, to perform database updates on a database resource that should be a slave.  This condition is to be avoided at all costs since the result is data loss or data corruption.